CN103440292B - Multimedia information retrieval method and system based on bit vectors - Google Patents

Abstract

The invention discloses a kind of multimedia information retrieval method and system based on bit vectors, described method includes: after extracting the characteristic of present multimedia information, obtains the high dimensional feature vector of the n dimension of present multimedia information；The intermediate vector of m dimension is obtained by projection matrix after being converted by the high dimensional feature vector that n ties up；Each element respective element with intermediate vector respectively of the threshold vector tieed up by m compares, and according to comparative result, intermediate vector is carried out binaryzation, obtains the bit vectors of the m dimension of present multimedia information；Wherein, m is less than n；According to the bit vectors obtained, finding out the bit vectors similar to this bit vectors in characteristics of the multimedia data base, the multimedia messages corresponding to bit vectors that will find out is as retrieval result output.This method ensure that original vector identification ability, after the bit vectors that high dimensional feature DUAL PROBLEMS OF VECTOR MAPPING is low-dimensional of multimedia messages so that recall precision based on bit vectors is higher, retrieval consumes less.

Description

Multimedia information retrieval method and system based on bit vectors

Technical field

The present invention relates to computer realm, particularly relate to a kind of multimedia information retrieval method based on bit vectors and be System.

Background technology

In recent years, along with the developing rapidly of multimedia technology and computer technology, large-scale multimedia messages increasingly How to occur in numerous research and application.In order to enable the information included in these numerous and jumbled data to obtain effectively Ground accesses and utilizes, and traditional text based retrieval technology cannot meet the demand that user is growing, based on content Retrieval technique just arise at the historic moment.

Content-based retrieval method needs first to extract multimedia characteristic and sets up property data base, then by right The retrieval conversion of multimedia messages is the retrieval of the neighbour to characteristic.For large scale multimedia information, its characteristic number According to being also large-scale.This needs exist for the suitable indexing means corresponding with characteristic and carrys out tissue characteristic data, accelerates The speed of retrieval.

But, the characteristic of multimedia messages is often the vector data of higher-dimension (being called for short high dimensional feature vector), tradition The Indexing Mechanism being adapted to low-dimensional data be difficult in adapt to the requirement in information retrieval based on contents, this most usually said higher-dimension The index dimension disaster phenomenon of data.To expend huge it is to say, realize the retrieval of multimedia messages based on high dimensional feature vector Big retrieval resource, consumption are very big, inefficiency.

For solving the problems referred to above, the method for prior art, such as similar sensitive hash (Similarity Sensitive Hash, SSH), local sensitivity Hash (Locality Sensitive Hash, LSH) method, by high dimensional feature vector is reflected Penetrate the bit vectors for low-dimensional, thus utilize Similarity Measures based on bit vectors and efficient index method to accelerate higher-dimension The retrieval rate of characteristic vector, thus improve the recall precision of multimedia messages.But, the method for prior art easily causes similar High dimensional feature vector (the most similar high dimensional feature vector) be mapped as dissimilar bit vectors, dissimilar high dimensional feature Vector (the most non-similar high dimensional feature vector) is mapped as similar bit vectors, when causing carrying out multimedia information retrieval, After the high dimensional feature DUAL PROBLEMS OF VECTOR MAPPING of multimedia messages is bit vectors, there is bigger erroneous matching rate so that original vector Identification ability declines.

Therefore, it is necessary to provide a kind of multimedia information retrieval method based on bit vectors, ensureing that original vector is known In the case of other ability, by the bit vectors that high dimensional feature DUAL PROBLEMS OF VECTOR MAPPING is low-dimensional of multimedia messages, so that based on bit The recall precision of the multimedia messages of vector is higher compared to the recall precision of multimedia messages based on high dimensional feature vector, subtracts Little retrieval consumes, and reduces the erroneous matching rate of the retrieval of multimedia messages based on bit vectors.

Summary of the invention

The defect existed for above-mentioned prior art, the invention provides the inspection of a kind of multimedia messages based on bit vectors Rope method and system, in order to, in the case of ensureing original vector identification ability, to reflect the high dimensional feature vector of multimedia messages After penetrating as the bit vectors of low-dimensional so that recall precision based on bit vectors is higher, retrieval consumes less.

According to an aspect of the invention, it is provided a kind of multimedia information retrieval method based on bit vectors, including:

After extracting the characteristic of present multimedia information, obtain the high dimensional feature of the n dimension of described present multimedia information Vector, is designated as X (x₁,x₂,...,x_n)；

By high dimensional feature vector X (x₁,x₂,...,x_n) by obtaining the intermediate vector W (w of m dimension after projection matrix P conversion₁, w₂,...,w_m)；

Each element respective element with described intermediate vector respectively of the threshold vector tieed up by m compares, according to comparing Result carries out binaryzation to described intermediate vector, obtains the bit vectors of the m dimension of described present multimedia information；Wherein, m is less than n；

According to the bit vectors obtained, find out in characteristics of the multimedia data base the bit similar to this bit vectors to Amount, the multimedia messages corresponding to bit vectors that will find out is as retrieval result output；

Wherein, described projection matrix P is the matrix of m × n, and meets following condition: in information bank storage each the most The high dimensional feature vector of the multimedia messages of classification, the most similar high dimensional feature vector vectorial spacing after P converts Expected value, minimum with the difference of the inhomogeneous high dimensional feature vector vectorial spacing expected value after P converts；

Described threshold vector meets following condition: for the high dimensional feature of each multimedia messages of storage in described information bank Vector, the most similar high dimensional feature vector is through P conversion and compares through described threshold vector, between vector after binaryzation Distance expected value, and compares through described threshold vector, after binaryzation through P conversion with inhomogeneous high dimensional feature vector The difference of vector spacing expected value is minimum.

It is preferred that before the characteristic of described extraction present multimedia information, also include:

Described projection matrix P is trained by the multimedia messages of storage in described information bank:

For the multimedia messages of storage in described information bank, by the higher-dimension of wherein any pair similar multimedia messages Characteristic vector, as a set element, stores in similar sample set；And

Using the high dimensional feature vector of wherein any pair inhomogeneous multimedia messages as a set element, storage is arrived In non-similar sample set；

Construct so that in equation below 1Minimum projection matrix P:

Wherein, Q is described similar sample set；R is described non-similar sample set；E{||PX-PX'||²| Q} represents institute State the high dimensional feature vector similar in Q vectorial spacing expected value after P converts；E{||PX-PX'||²| R} represents described The inhomogeneous high dimensional feature vector vectorial spacing expected value after P converts in R；α is the weights set.

Construct described in it is preferred that so that in equation below 1Minimum projection matrix P, specifically includes:

Ask for matrix ∑_GM minimum n tie up matrix characteristic vector；Wherein,Described ∑_QSuch as formula 2 shown, described ∑s_RAs shown in Equation 3:

∑_Q=E{ (X-X') (X-X')^T| Q} (formula 2)

In described formula 2, E{ (X-X') (X-X')^T| Q} represents the covariance between high dimensional feature vector similar in described Q The average of matrix；

∑_R=E{ (X-X') (X-X')^T| R} (formula 3)

In described formula 3, E{ (X-X') (X-X')^T| R} represents the association side in described R between inhomogeneous high dimensional feature vector The average of difference matrix；

Tieed up matrix characteristic vector by m the n asked for, constitute the projection matrix P of m × n.

It is preferred that described by described information bank in after the multimedia messages of storage trains described projection matrix P, also Including:

Calculate the m dimensional vector minimum so that L in equation below 4, be designated as U (u₁,u₂,...,u_m), and as described threshold value Vector:

L=E{sign (PX+U)^Tsign(PX'+U)|R}-αE{sign(PX+U)^TSign (PX'+U) | Q} (formula 4)

Wherein, E{sign (PX+U)^TSign (PX'+U) | Q} represents that high dimensional feature vector similar in described Q becomes through P Change and after described threshold vector compares and determines sign symbol, the average of the distance between the symbolic vector obtained；E{sign (PX+U)^TSign (PX'+U) | R} represents that in described R, inhomogeneous high dimensional feature vector is through P conversion and through described threshold value After vector compares and determines sign symbol, the average of the distance between the symbolic vector obtained.

Or, described by described information bank in storage multimedia messages train described projection matrix P after, also wrap Include:

Calculate the m dimensional vector minimum so that L in equation below 4, be designated as U (u₁,u₂,...,u_m):

L=E{sign (PX+U)^Tsign(PX'+U)|R}-αE{sign(PX+U)^TSign (PX'+U) | Q} (formula 4)

Wherein, E{sign (PX+U)^TSign (PX'+U) | Q} represents that high dimensional feature vector similar in described Q becomes through P Change and after described threshold vector compares and determines sign symbol, the average of the distance between the symbolic vector obtained；E{sign (PX+U)^TSign (PX'+U) | R} represents that in described R, inhomogeneous high dimensional feature vector is through P conversion and through described threshold value After vector compares and determines sign symbol, the average of the distance between the symbolic vector obtained；

Afterwards, to U (u₁,u₂,...,u_m) be optimized after, obtain described threshold vector:

Element u for described threshold vector U_i, utilize equation below 5 and formula 6, ask for so that FN (u_i)+α×FP (u_i) minimum u_iValue, as the u after optimizing_iValue；

FN(u_i)=Pr (min{z, z'} >=u_i or max{z,z'}<u_i| Q) (formula 5)

FP(u_i)=Pr (min{z, z'} <u_i≤ max{z, z'} | R) (formula 6)

In described formula 5, (min{z, z'} >=u_i or max{z,z'}<u_i| Q) in z and z' represent in described Q arbitrarily A pair similar high dimensional feature vector X and X' in one set element obtains after converting respectively through described projection matrix P The i-th element of vector, Pr (min{z, z'} >=u_i or max{z,z'}<u_i| Q) represent for the set element in described Q, u_iMeet following condition: min{z, z'} >=u_i or max{z,z'}<u_iProbability；

In described formula 6, (min{z, z'} <u_i≤ max{z, z'} | R) in z and z' represent in described R any one collection Close the vector obtained after a pair inhomogeneous high dimensional feature vector X and X' in element converts respectively through described projection matrix P I-th element, Pr (min{z, z'} <u_i≤ max{z, z'} | R) represent for the set element in described R, u_iMeet as follows Condition: min{z, z'} <u_iThe probability of≤max{z, z'}.

Calculate the m dimensional vector minimum so that following L described in it is preferred that, specifically include:

Ask for so that the u of following expression 7 minimum_iValue；Wherein, i is the natural number of 1～m；

E{sign((P_i ^TX+u_i)(P_i ^TX'+u_i))|R}-αE{sign((P_i ^TX+u_i)^T(P_i ^TX'+u_i)) | Q} (expression formula 7)

Wherein, P_i ^TThe i-th row vector for described projection matrix P；u_iFor U (u₁,u₂,...,u_m) i-th element；

And the u that will obtain₁～u_mForm described m dimensional vector.

According to another aspect of the present invention, additionally provide a kind of Multimedia information retrieval system based on bit vectors, Including:

Bit vectors modular converter, after the characteristic extracting present multimedia information, obtains described current many matchmakers The high dimensional feature vector of the n dimension of body information, is designated as X (x₁,x₂,...,x_n)；By high dimensional feature vector X (x₁,x₂,...,x_n) pass through Intermediate vector W (the w of m dimension is obtained after projection matrix P conversion₁,w₂,...,w_mAfter), each element of the threshold vector tieed up by m is respectively Compare with the respective element of described intermediate vector, according to comparative result, described intermediate vector is carried out binaryzation, obtain institute State the bit vectors of the m dimension of present multimedia information；Wherein, m is less than n；

Retrieval module, the bit vectors of the present multimedia information for obtaining according to described bit vectors modular converter, The bit vectors similar to this bit vectors is found out, corresponding to the bit vectors that will find out in characteristics of the multimedia data base Multimedia messages as retrieval result output；

Wherein, described projection matrix P is the matrix of m × n, and meets following condition: each many for store in information bank The high dimensional feature vector of media information, the most similar high dimensional feature vector vectorial spacing expected value after P converts, with The difference of the inhomogeneous high dimensional feature vector vectorial spacing expected value after P converts is minimum；

Described threshold vector meets following condition: for each classified multimedia messages of storage in described information bank High dimensional feature vector, the most similar high dimensional feature vector is through P conversion and compares through described threshold vector, after binaryzation Vectorial spacing expected value, with inhomogeneous high dimensional feature vector through P conversion and compare through described threshold vector, two The difference of the vectorial spacing expected value after value is minimum.

It is preferred that described bit vectors modular converter specifically includes:

High dimensional feature vector determination unit, after extracting the characteristic of present multimedia information, obtain described currently The high dimensional feature vector of the n dimension of multimedia messages, is designated as X (x₁,x₂,...,x_n)；

Intermediate vector computing unit, for the high dimensional feature vector X obtained by described high dimensional feature vector determination unit (x₁,x₂,...,x_n) by obtaining the intermediate vector W (w of m dimension after projection matrix P conversion₁,w₂,...,w_m)；

Threshold value comparing unit, each element of the threshold vector for being tieed up by m obtains with described intermediate vector computing unit respectively To the respective element of intermediate vector compare, according to comparative result, described intermediate vector is carried out binaryzation, obtains described The bit vectors of the m dimension of present multimedia information；Wherein, m is less than n.

Further, described Multimedia information retrieval system based on bit vectors, also include:

Projection matrix builds module, for training described projection square by the multimedia messages of storage in described information bank Battle array P: for the multimedia messages of storage in described information bank, by the high dimensional feature of wherein any pair similar multimedia messages Vector, as a set element, stores in similar sample set；And by wherein any pair inhomogeneous multimedia messages High dimensional feature vector as a set element, store in non-similar sample set；Construct so that in equation below 1 Minimum projection matrix P:

Wherein, Q is described similar sample set；R is described non-similar sample set；E{||PX-PX'||²| Q} represents institute State the high dimensional feature vector similar in Q vectorial spacing expected value after P converts；E{||PX-PX'||²| R} represents described The inhomogeneous high dimensional feature vector vectorial spacing expected value after P converts in R；α is the weights set；

First threshold vector determines module, for calculating the m dimensional vector minimum so that L in equation below 4, is designated as U (u₁,u₂,...,u_m), and as described threshold vector:

L=E{sign (PX+U)^Tsign(PX'+U)|R}-αE{sign(PX+U)^TSign (PX'+U) | Q} (formula 4)

Wherein, E{sign (PX+U)^TSign (PX'+U) | Q} represents that high dimensional feature vector similar in described Q becomes through P Change and after described threshold vector compares and determines sign symbol, the average of the distance between the symbolic vector obtained；E{sign (PX+U)^TSign (PX'+U) | R} represents that in described R, inhomogeneous high dimensional feature vector is through P conversion and through described threshold value After vector compares and determines sign symbol, the average of the distance between the symbolic vector obtained.

It is preferred that described first threshold vector determines that module specifically includes:

Minimum calculation unit, for asking for so that the u of following expression 7 minimum_iValue；Wherein, i is the nature of 1～m Number；

E{sign((P_i ^TX+u_i)(P_i ^TX'+u_i))|R}-αE{sign((P_i ^TX+u_i)^T(P_i ^TX'+u_i)) | Q} (expression formula 7)

Wherein, P_i ^TThe i-th row vector for described projection matrix P；u_iFor U (u₁,u₂,...,u_m) i-th element；

Vector component units, for the u described minimum calculation unit obtained₁～u_mForm described m dimensional vector U (u₁, u₂,...,u_m), as described threshold vector.

Further, described Multimedia information retrieval system based on bit vectors, also include:

Projection matrix builds module, for training described projection square by the multimedia messages of storage in described information bank Battle array P: for the multimedia messages of storage in described information bank, by the high dimensional feature of wherein any pair similar multimedia messages Vector, as a set element, stores in similar sample set；And by wherein any pair inhomogeneous multimedia messages High dimensional feature vector as a set element, store in non-similar sample set；Construct so that in equation below 1 Minimum projection matrix P:

Wherein, Q is described similar sample set；R is described non-similar sample set；E{||PX-PX'||²| Q} represents institute State the high dimensional feature vector similar in Q vectorial spacing expected value after P converts；E{||PX-PX'||²| R} represents described The inhomogeneous high dimensional feature vector vectorial spacing expected value after P converts in R；α is the weights set；

Second Threshold vector determines module, for calculating the m dimensional vector minimum so that L in equation below 4, is designated as U (u₁,u₂,...,u_m):

L=E{sign (PX+U)^Tsign(PX'+U)|R}-αE{sign(PX+U)^TSign (PX'+U) | Q} (formula 4)

Wherein, E{sign (PX+U)^TSign (PX'+U) | Q} represents that high dimensional feature vector similar in described Q becomes through P Change and after described threshold vector compares and determines sign symbol, the average of the distance between the symbolic vector obtained；E{sign (PX+U)^TSign (PX'+U) | R} represents that in described R, inhomogeneous high dimensional feature vector is through P conversion and through described threshold value After vector compares and determines sign symbol, the average of the distance between the symbolic vector obtained；

Second Threshold vector determines that module is to U (u₁,u₂,...,u_m) be optimized after, obtain described threshold vector.

It is preferred that described Second Threshold vector determines that module specifically includes:

Minimum calculation unit, for asking for so that the u of following expression 7 minimum_iValue；Wherein, i is the nature of 1～m Number；

E{sign((P_i ^TX+u_i)(P_i ^TX'+u_i))|R}-αE{sign((P_i ^TX+u_i)^T(P_i ^TX'+u_i)) | Q} (expression formula 7)

Wherein, P_i ^TThe i-th row vector for described projection matrix P；u_iFor U (u₁,u₂,...,u_m) i-th element；

Vector optimization unit, for U (u₁,u₂,...,u_m) element u_iIt is optimized: for described threshold vector U's Element u_i, utilize equation below 5 and formula 6, ask for so that FN (u_i)+α×FP(u_i) minimum u_iValue, as the u after optimizing_i Value；

FN(u_i)=Pr (min{z, z'} >=u_i or max{z,z'}<u_i| Q) (formula 5)

FP(u_i)=Pr (min{z, z'} <u_i≤ max{z, z'} | R) (formula 6)

In described formula 5, (min{z, z'} >=u_i or max{z,z'}<u_i| Q) in z and z' represent in described Q arbitrarily A pair similar high dimensional feature vector X and X' in one set element obtains after converting respectively through described projection matrix P The i-th element of vector, Pr (min{z, z'} >=u_i or max{z,z'}<u_i| Q) represent for the set element in described Q, u_iMeet following condition: min{z, z'} >=u_i or max{z,z'}<u_iProbability；

In described formula 6, (min{z, z'} <u_i≤ max{z, z'} | R) in z and z' represent in described R any one collection Close the vector obtained after a pair inhomogeneous high dimensional feature vector X and X' in element converts respectively through described projection matrix P I-th element, Pr (min{z, z'} <u_i≤ max{z, z'} | R) represent for the set element in described R, u_iMeet as follows Condition: min{z, z'} <u_iThe probability of≤max{z, z'}；

Vector component units, for by the u after described vector optimization unit optimization₁～u_mForm described threshold vector.

Specifically include it is preferred that described projection matrix builds module:

Minimum matrix characteristic vector computing unit, is used for asking for matrix ∑_GM minimum n tie up matrix characteristic vector；Its In,Described ∑_QAs shown in Equation 2, described ∑_RAs shown in Equation 3:

∑_Q=E{ (X-X') (X-X')^T| Q} (formula 2)

In described formula 2, E{ (X-X') (X-X')^T| Q} represents the covariance between high dimensional feature vector similar in described Q The average of matrix；

Σ_R=E{ (X-X') (X-X')^T| R} (formula 3)

In described formula 3, E{ (X-X') (X-X')^T| R} represents the association side in described R between inhomogeneous high dimensional feature vector The average of difference matrix；

Projection matrix determines unit, for being tieed up matrix characteristic vector by m the n asked for, constitutes the projection matrix P of m × n.

In technical scheme, have after being converted into bit vectors due to the high dimensional feature vector of present multimedia information Effect discrete between gathering, class in having class, thus ensure that original vector identification ability；So, application ripe based on low-dimensional The retrieval technique of bit vectors, it is possible to achieve compared to the higher recall precision of retrieval technique based on high dimensional feature vector, Less retrieval consumption, and the retrieval result that the retrieval of multimedia messages based on bit vectors is drawn more is as the criterion Really, the erroneous matching rate of retrieval is reduced.

Accompanying drawing explanation

Fig. 1 a is the method training projection matrix according to the multimedia messages of storage in information bank of the embodiment of the present invention Flow chart；

Fig. 1 b be the embodiment of the present invention according to Σ_GConstruct the flow chart of the concrete grammar of projection matrix；

Fig. 2 is the flow chart of the multimedia information retrieval method based on bit vectors of the embodiment of the present invention；

Fig. 3 a is a kind of internal structure frame of the Multimedia information retrieval system based on bit vectors of the embodiment of the present invention Figure；

Fig. 3 b is the another kind of internal structure of the Multimedia information retrieval system based on bit vectors of the embodiment of the present invention Block diagram；

Fig. 4 is the method flow diagram carrying out multimedia information retrieval according to bit vectors of the embodiment of the present invention.

Detailed description of the invention

Below with reference to accompanying drawing, technical scheme is carried out clear, complete description, it is clear that described enforcement Example is only a part of embodiment of the present invention rather than whole embodiments.Based on the embodiment in the present invention, this area is general All other embodiments that logical technical staff is obtained on the premise of not making creative work, broadly fall into the present invention and are protected The scope protected.

The term such as " module " used in this application, " system " is intended to include the entity relevant to computer, such as but does not limits In hardware, firmware, combination thereof, software or executory software.Such as, module it may be that it is not limited to: process Process, processor, object, executable program, the thread of execution, program and/or the computer run on device.For example, meter Application program and this calculating equipment of running on calculation equipment can be modules.One or more modules may be located at executory In one process and/or thread, a module can also be positioned on a computer and/or be distributed in two or the calculating of more multiple stage Between machine.

In technical scheme, construct a mapping function, use this mapping function can by high dimensional feature to Amount is mapped as the bit vectors of low-dimensional, and this mapping function can also ensure: original similar high dimensional feature vector, Jing Guoying The bit vectors obtained after penetrating is even more like；Original high dimensional feature vector is dissimilar, and the bit vectors obtained after mapping is more Add dissmilarity；It is to say, through the mapping of this mapping function, have after original high dimensional feature vector is converted into bit vectors Effect discrete between gathering, class in class, thus ensure original vector identification ability；Afterwards, the ratio based on low-dimensional that application is ripe The retrieval technique of special vector, it is achieved compared to the higher recall precision of retrieval technique based on high dimensional feature vector and less Retrieval consumes.

Describe technical scheme below in conjunction with the accompanying drawings in detail.The embodiment of the present invention is extracting present multimedia information Characteristic, before the retrieval of the characteristic carrying out present multimedia information, needing first to construct can be by current many matchmaker The mapping function of the binaryzation that high dimensional feature DUAL PROBLEMS OF VECTOR MAPPING the is low-dimensional vector of the n dimension of body information, is designated as:

Y=sign (PX+U),

Wherein, P is the projection matrix of m × n；U is the threshold vector of m dimension, is designated as U (u₁,u₂,...,u_m)；X is the height of n dimension Dimensional feature vector, is designated as X (x₁,x₂,...,x_n), and each element in X is real number value；Sign (PX+U) represents amount of orientation PX+U Symbol (sign), obtain the symbolic vector (element of symbolic vector be-1 or+1) of binaryzation, the element of even PX+U Symbol is negative sign, and the respective element in then symbol vector is-1, if the symbol of the element of PX+U is positive sign, in then symbol vector Respective element is+1；Y is the symbolic vector of the binaryzation of the m dimension obtained after the symbol of amount of orientation PX+U, is designated as Y (y₁,y₂,..., y_m)；It is true that each element in symbolic vector can represent with bit, such as, symbol is that the element of negative sign can be with bit 0 table Showing, symbol is that the element of positive sign can represent with bit 1, thus obtains corresponding bit vectors.

In follow-up, with the high dimensional feature vector X (x of n dimension₁,x₂,...,x_n) be column vector to construct mapping function, and Mapping function according to structure obtains the column vector of m dimension, the i.e. bit vectors of m dimension；Those skilled in the art can basis Technical scheme disclosed in the embodiment of the present invention, the high dimensional feature vector X (x that easy realization is tieed up with n₁,x₂,...,x_n) it is Row vector constructs mapping function, and then maps the technical scheme of the bit vectors of the row vector obtaining m dimension；Therefore, no matter with The high dimensional feature vector X (x of row vector or column vector₁,x₂,...,x_n) construct mapping function and then map the ratio obtaining m dimension Method or the design of special vector all should be within protection scope of the present invention.

Specifically, projection matrix P, and training can be trained according to the classified multimedia messages of storage in information bank The projection matrix P of the m × n gone out meets following condition: for the higher-dimension of each classified multimedia messages of storage in information bank Characteristic vector, the most similar high dimensional feature vector vectorial spacing expected value after P converts, special with inhomogeneous higher-dimension The difference levying the vector vectorial spacing expected value after P converts is minimum.As shown in Figure 1a, many according to information bank stores Media information trains the method for projection matrix P, comprises the steps:

S101: for the multimedia messages of storage in information bank, by the height of wherein any pair similar multimedia messages Dimensional feature vector, as a set element, stores in similar sample set；And by wherein any pair inhomogeneous many matchmaker The high dimensional feature vector of body information, as a set element, stores in non-similar sample set.

Specifically, for the multimedia messages of storage in information bank, according between the high dimensional feature vector of multimedia messages Similarity, has pre-build the similar sample set comprising similar high dimensional feature vector, has been designated as Q, and comprises inhomogeneous The non-similar sample set of high dimensional feature vector, is designated as R.

S102: construct so that in equation below 1Minimum projection matrix P:

Above-mentioned formula 1 is predefined object function；Wherein, Q is similar sample set；R is non-similar sample set Close；{||PX-PX'||²| a pair similar high dimensional feature vector that X and X' in Q} represents in Q in any one set element； {||PX-PX'||²| a pair inhomogeneous high dimensional feature vector that X and X' in R} represents in R in any one set element； PX-PX' represents the distance between the vector that high dimensional feature vector X and X' obtains after P converts；||PX-PX'||²Represent height The covariance of the distance between the vector that dimensional feature vector X and X' obtains after P converts；

E{||PX-PX'||²| Q} represents the vectorial vectorial spacing expectation after P converts of high dimensional feature similar in Q Value, the average of the covariance of the vectorial vectorial spacing after P converts of high dimensional feature i.e. representing similar in Q；E{||PX- PX'||²| R} represents the vector of inhomogeneous high dimensional feature in R vectorial spacing expected value after P converts, and i.e. represents in R not The average of the covariance of the similar high dimensional feature vector vectorial spacing after P converts；α is the weights set, and value is 1 ～0.5；α is specially the power of the tolerance ratio of similar high dimensional feature vector spacing and non-similar high dimensional feature vector spacing Value, weights are the biggest, and similar high dimensional feature vector distance metric weights is the biggest, and similar high dimensional feature vector becomes through projection matrix P In changing rear class, aggregation extent is the highest, and in other words, inhomogeneity high dimensional feature vector distance metric weights is the least, inhomogeneity high dimensional feature Vector after projection matrix P converts between class dispersion degree the highest.

Specifically, according to the knowledge of linear algebra, it can be deduced that:

E{||PX-PX'||²| Q}=tr{P ∑_QP^T(formula 8)

E{||PX-PX'||²| R}=tr{P ∑_RP^T(formula 9)

Wherein, P^TRepresent the transposed matrix seeking P；tr{P∑_QP^TRepresent and seek matrix P ∑_QP^TMark, tr{P ∑_RP^TRepresent Seek matrix P ∑_RP^TMark；∑_QAs shown in Equation 2, ∑_RAs shown in Equation 3:

∑_Q=E{ (X-X') (X-X')^T| Q} (formula 2)

In formula 2, { (X-X') (X-X')^T| represent in Q in any one set element a pair of X and X' in Q} is similar High dimensional feature vector, wherein, (X-X')^TRepresent the transposed vector asking for (X-X')；E{(X-X')(X-X')^T| Q} represents in Q The average of the similar covariance matrix between high dimensional feature vector, specifically represents the association between high dimensional feature vector similar in Q Each element of variance matrix is averaged；

∑_R=E{ (X-X') (X-X')^T| R} (formula 3)

In formula 3, E{ (X-X') (X-X')^T| X and X' in R} represent in R in any one set element a pair is not Similar high dimensional feature vector, E{ (X-X') (X-X')^T| R} represents the covariance square in R between inhomogeneous high dimensional feature vector The average of battle array, each element of the covariance matrix between high dimensional feature vector inhomogeneous in R is averaged by concrete expression.

So, according to formula 8 and formula 9, above-mentioned formula 1 can be converted into formula 10:

Further, useRepresent and ask for Σ_RInverse matrixAfter, rightEvolution) it is multiplied by formula 10 The right expression formula after, the expression formula obtained is multiplied by againExpression is asked forTransposed matrix) after, Make tr{R Σ_RR^TIt is converted into constant, make tr{P Σ_QP^TBe converted to such as the expression formula on the right in formula 11:

Formula 11 representsIt is proportional to

And,

$\begin{array}{c}tr\left\{{\mathrm{P\&Sigma;}}_R^{-1/2}{\mathrm{\&Sigma;}}_Q{\mathrm{\&Sigma;}}_R^{-T/2}{P}^T\right\}\\ =tr\left\{{\mathrm{P\&Sigma;}}_Q{\mathrm{\&Sigma;}}_R^{-1}{P}^T\right\}\\ =tr\left\{{\mathrm{P\&Sigma;}}_G{P}^T\right\}\end{array}$

Wherein,

As such, it is possible to according to Σ_G, construct so that in formula 1Minimum projection matrix P, the flow chart of its concrete grammar As shown in Figure 1 b, comprise the steps:

S111: ask for Σ_GM minimum n tie up matrix characteristic vector.

Specifically, ∑_GIt is a positive semidefinite symmetrical matrix, matrix ∑ can be asked for according to linear algebra knowledge_GM Little characteristic vector, i.e. obtains m minimum n and ties up matrix characteristic vector.

S112: tieed up matrix characteristic vector by m the n asked for, constitute the projection matrix P of m × n.

Specifically, m the n asked for tie up matrix characteristic vector, constitute the orthogonal matrix of m × n, i.e. projection matrix P；This throwing Shadow matrix P is so that in formula 1Obtain minima.

After the multimedia messages of storage trains projection matrix P in by information bank, threshold vector U can be calculated, And threshold vector U meets following condition: for the high dimensional feature vector of each multimedia messages of storage in information bank, the most similar High dimensional feature vector through P conversion and compare through threshold vector, vectorial spacing expected value after binaryzation, from different The high dimensional feature vector of class through P conversion and compare through threshold vector, the difference of vectorial spacing expected value after binaryzation Minimum.

Wherein, calculate threshold vector U, specially calculate the m dimensional vector minimum so that the L in equation below 4, as Threshold vector U:

L=E{sign (PX+U)^Tsign(PX'+U)|R}-αE{sign(PX+U)^TSign (PX'+U) | Q} (formula 4)

Wherein, E{sign (PX+U)^TSign (PX'+U) | Q} represent similar in Q high dimensional feature vector through P conversion and After threshold vector U compares and determines sign symbol, the average of the distance between the symbolic vector obtained；E{sign(PX+U)^TSign (PX'+U) | R} represents in R that inhomogeneous high dimensional feature vector just determines through P conversion comparing through threshold vector U After minus symbol, the average of the distance between the symbolic vector obtained；Wherein, the distance between symbolic vector reflect this symbol to Amount carries out the distance between the bit vectors after binaryzation.

Further, formula 4 is converted:

$\begin{array}{c}L=E\left\{sign{(PX+U)}^{T}sign({\mathrm{PX}}^{\&prime;}+U)\right|R\}-\mathrm{\&alpha;}E\{sign{(PX+U)}^{T}sign({\mathrm{PX}}^{\&prime;}+U)\left|Q\right\}\\ =\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\left\{E\right\{sign(P_i^{T}X+u_i)sign(P_i^T{X}^{\&prime;}+u_i)\left|R\right\}-\mathrm{\&alpha;}E\left\{sign(P_i^{T}X+u_i)sign(P_i^T{X}^{\&prime;}+u_i)\right|Q\left\}\right\}\\ =\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}\left\{E\right\{sign\left(\right(P_i^{T}X+u_i\left)\right(P_i^T{X}^{\&prime;}+u_i\left)\right)\left|R\right\}-\mathrm{\&alpha;}E\left\{sign\left(\right(P_i^{T}X+u_i\left)\right(P_i^T{X}^{\&prime;}+u_i\left)\right)\right|Q\left\}\right\}\end{array}$

Wherein, P_i ^TRepresent i-th row vector of projection matrix P；u_iFor U (u₁,u₂,...,u_m) i-th element；I is 1～m Natural number.

As such, it is possible to by asking for so that m minimum for L ties up threshold vector, be converted into m independent asking for so that being expressed as The u of formula 7 minimum_iValue:

E{sign((P_i ^TX+u_i)(P_i ^TX'+u_i))|R}-αE{sign((P_i ^TX+u_i)^T(P_i ^TX'+u_i)) | Q} (expression formula 7)

Calculating so that the u of expression formula 7 minimum_iAfter value, the u that will obtain₁～u_mComposition m dimensional vector, can will obtain M dimensional vector is as threshold vector U；As a kind of more excellent embodiment, can also continue to obtain by u₁～u_mComposition m dimension to Amount is optimized, and the m dimensional vector after optimizing is as final threshold vector U:

Specifically, for the element u calculated_i, utilize equation below 5 and formula 6, ask for so that FN (u_i)+α×FP (u_i) minimum u_iValue, as the u after optimizing_iValue:

FN(u_i)=Pr (min{z, z'} >=u_i or max{z,z'}<u_i| Q) (formula 5)

FP(u_i)=Pr (min{z, z'} <u_i≤ max{z, z'} | R) (formula 6)

Wherein, z=P_i ^TX and z'=P_i ^TX'；Min{z, z'} represent the minima asked in two element z and z', max { z, z'} represent the maximum asked in two element z and z'；

In formula 5, (min{z, z'} >=u_i or max{z,z'}<u_i| Q) in z and z' represent in Q any one set The i-th unit of the vector that a pair similar high dimensional feature vector X and X' in element obtains after converting respectively through projection matrix P Element, Pr (min{z, z'} >=u_i or max{z,z'}<u_i| Q) represent for the set element in Q, u_iMeet following condition: min {z,z'}≥u_i or max{z,z'}<u_iProbability；

In formula 6, (min{z, z'} <u_i≤ max{z, z'} | R) in z and z' represent in R in any one set element A pair inhomogeneous high dimensional feature vector X and X' respectively through projection matrix P convert after obtain vector i-th element, Pr(min{z,z'}<u_i≤ max{z, z'} | R) represent for the set element in R, u_iMeet following condition: min{z, z'} <u_i The probability of≤max{z, z'}；

U after optimizing₁～u_mComposition m dimensional vector is as final threshold vector U.

Due to Section 1 the E{sign ((P in expression formula 7_i ^TX+u_i)(P_i ^TX'+u_i)) | the value of R} is proportional to FP (u_i), the Binomial E{sign ((P_i ^TX+u_i)^T(P_i ^TX'+u_i)) | the value of Q} is proportional to FN (u_i), and according to mathematical statistics knowledge, can hold very much Easy estimates according to classified multimedia messages, therefore, it can by asking for so that FN (u_i)+α×FP(u_i) minimum u_iValue, determines the final threshold vector U after optimization quickly and accurately.

According to above-mentioned method, after constructing projection matrix P and threshold vector U, can construct and present multimedia is believed The mapping function Y=sign (PX+U) of the binaryzation that high dimensional feature DUAL PROBLEMS OF VECTOR MAPPING the is low-dimensional vector of the n dimension of breath.And, pass through The mapping of this mapping function, effect discrete between gathering, class in there is after original high dimensional feature vector is converted into bit vectors class Really, thus ensure that the identification ability of original vector.

The mapping function using above-mentioned structure can be by the binaryzation vector that high dimensional feature DUAL PROBLEMS OF VECTOR MAPPING is low-dimensional, Jin Erjin Row multimedia information retrieval based on bit vectors, the flow chart of its method is as in figure 2 it is shown, comprise the steps:

S201: after extracting the characteristic of present multimedia information, obtains the high dimensional feature of the n dimension of present multimedia information Vector X (x₁,x₂,...,x_n)。

S202: by X (x₁,x₂,...,x_n) by obtaining the intermediate vector W (w of m dimension after projection matrix P conversion₁,w₂,..., w_m)。

Specifically, can according to structure mapping function time, constructed by the projection matrix P that goes out high dimensional feature vector X that n is tieed up (x₁,x₂,...,x_n) convert, obtain intermediate vector PX of m dimension, be designated as W (w₁,w₂,...,w_m)。

Each element respective element with intermediate vector respectively of S203: the threshold vector tieed up by m compares, according to comparing Result carries out binaryzation to intermediate vector, obtains the bit vectors of the m dimension of present multimedia information；Wherein, m is less than n.

Specifically, can be according to when constructing mapping function, the threshold vector U of the m dimension calculated, by U (u₁,u₂,..., u_m) each element respectively with intermediate vector W (w₁,w₂,...,w_m) respective element compare, according to comparative result to centre Vector carries out binaryzation, obtains the bit vectors of the m dimension of present multimedia information.

Wherein it is possible to intermediate vector is carried out binaryzation according to mapping function: ask for W+U, i.e., after PX+U, ask for sign (PX+U) and after obtaining symbolic vector, each element of symbolic vector is represented with bit (0 or 1), obtain corresponding bit to Amount.So, owing to m is less than n, after intermediate vector is carried out binaryzation, it is achieved that the higher-dimension tieed up by the n of present multimedia information is special Levy the bit vectors that DUAL PROBLEMS OF VECTOR MAPPING is low-dimensional (m dimension).

S204: according to the bit vectors obtained, finds out similar to this bit vectors in characteristics of the multimedia data base Bit vectors, the multimedia messages corresponding to bit vectors that will find out is as retrieval result output.

Specifically, can be according to existing multimedia information retrieval method based on bit vectors (as shown in figure 4 below Method), carry out the retrieval of multimedia messages based on bit vectors, to obtain retrieving result.

The embodiment of the present invention additionally provides a kind of Multimedia information retrieval system based on bit vectors, its internal structure frame Figure, as shown in Fig. 3 a or 3b, specifically includes: bit vectors modular converter 301 and retrieval module 302.

Bit vectors modular converter 301, after the characteristic extracting present multimedia information, obtains present multimedia The high dimensional feature vector of the n dimension of information, is designated as X (x₁,x₂,...,x_n)；By high dimensional feature vector X (x₁,x₂,...,x_n) by throwing Intermediate vector W (the w of m dimension is obtained after shadow matrix P conversion₁,w₂,...,w_mAfter), by m tie up threshold vector each element respectively with The respective element of intermediate vector compares, and according to comparative result, intermediate vector is carried out binaryzation, obtains present multimedia letter The bit vectors of the m dimension of breath；Wherein, m is less than n.

The bit vectors of the retrieval module 302 present multimedia information for obtaining according to bit vectors modular converter 301, The bit vectors similar to this bit vectors is found out, corresponding to the bit vectors that will find out in characteristics of the multimedia data base Multimedia messages as retrieval result output.

Wherein, projection matrix P is the matrix of m × n, and meets following condition: classify for each of storage in information bank The high dimensional feature vector of multimedia messages, the vectorial spacing expectation after P converts of the most similar high dimensional feature vector Value, minimum with the difference of the inhomogeneous high dimensional feature vector vectorial spacing expected value after P converts.

Threshold vector meets following condition: for the high dimensional feature vector of each multimedia messages of storage in information bank, its In similar high dimensional feature vector through P conversion and compare through described threshold vector, vectorial spacing expectation after binaryzation Value, with inhomogeneous high dimensional feature vector through P conversion and compare through threshold vector, vectorial spacing phase after binaryzation The difference of prestige value is minimum.

Above-mentioned bit vectors modular converter 301 specifically includes: high dimensional feature vector determination unit 311, intermediate vector meter Calculate unit 312 and threshold value comparing unit 313.

High dimensional feature vector determination unit 311, after the characteristic extracting present multimedia information, obtains the most The high dimensional feature vector of the n dimension of media information, is designated as X (x₁,x₂,...,x_n)。

Intermediate vector computing unit 312 is for the high dimensional feature vector X obtained by high dimensional feature vector determination unit 311 (x₁,x₂,...,x_n) by obtaining the intermediate vector W (w of m dimension after projection matrix P conversion₁,w₂,...,w_m)。

Each element of the threshold value comparing unit 313 threshold vector for being tieed up by m obtains with intermediate vector computing unit 312 respectively To the respective element of intermediate vector compare, according to comparative result, intermediate vector is carried out binaryzation, obtains current many matchmakers The bit vectors of the m dimension of body information；Wherein, m is less than n.

Further, Multimedia information retrieval system based on bit vectors, also include: projection matrix builds module 303.

Projection matrix builds module 303 for training projection matrix P by the multimedia messages of storage in information bank: right The multimedia messages of storage in information bank, using the high dimensional feature vector of wherein any pair similar multimedia messages as one Individual set element, stores in similar sample set；And by the high dimensional feature of wherein any pair inhomogeneous multimedia messages Vector, as a set element, stores in non-similar sample set；Construct so that in equation below 1Minimum projection Matrix P:

Wherein, Q is described similar sample set；R is described non-similar sample set；E{||PX-PX'||²| Q} represents institute State the high dimensional feature vector similar in Q vectorial spacing expected value after P converts；E{||PX-PX'||²| R} represents described The inhomogeneous high dimensional feature vector vectorial spacing expected value after P converts in R；α is the weights set.

Above-mentioned shadow matrix builds module 303 and specifically includes: minimum matrix characteristic vector computing unit 331 and projection matrix Determine unit 332.

Minimum matrix characteristic vector computing unit 331 is used for asking for matrix ∑_GM minimum n tie up matrix characteristic vector； Wherein,Described ∑_QAs shown in Equation 2, described ∑_RAs shown in Equation 3:

∑_Q=E{ (X-X') (X-X')^T| Q} (formula 2)

In described formula 2, E{ (X-X') (X-X')^T| Q} represents the covariance between high dimensional feature vector similar in described Q The average of matrix；

∑_R=E{ (X-X') (X-X')^T| R} (formula 3)

In described formula 3, E{ (X-X') (X-X')^T| R} represents the association side in described R between inhomogeneous high dimensional feature vector The average of difference matrix.

Projection matrix determines the height of the unit 332 m the n dimension for being asked for by minimum matrix characteristic vector computing unit 331 Dimensional feature vector, constitutes the projection matrix P of m × n.

Further, Multimedia information retrieval system based on bit vectors, also include: first threshold vector determines module 304 (as shown in Figure 3 a), or Second Threshold vector determines module 305 (as shown in Figure 3 b).

First threshold vector determines that module 304, for calculating the m dimensional vector minimum so that L in equation below 4, is designated as U (u₁,u₂,...,u_m), and as described threshold vector:

L=E{sign (PX+U)^Tsign(PX'+U)|R}-αE{sign(PX+U)^TSign (PX'+U) | Q} (formula 4)

Wherein, E{sign (PX+U)^TSign (PX'+U) | Q} represents that high dimensional feature vector similar in described Q becomes through P Change and after described threshold vector compares and determines sign symbol, the average of the distance between the symbolic vector obtained；E{sign (PX+U)^TSign (PX'+U) | R} represents that in described R, inhomogeneous high dimensional feature vector is through P conversion and through described threshold value After vector compares and determines sign symbol, the average of the distance between the symbolic vector obtained.

Above-mentioned first threshold vector determines that module 304 specifically includes: minimum calculation unit 341 and vector component units 342。

Minimum calculation unit 341 is for asking for so that the u of following expression 7 minimum_iValue；Wherein, i is the nature of 1～m Number；

E{sign((P_i ^TX+u_i)(P_i ^TX'+u_i))|R}-αE{sign((P_i ^TX+u_i)^T(P_i ^TX'+u_i)) | Q} (expression formula 7)

Wherein, P_i ^TThe i-th row vector for described projection matrix P；u_iFor U (u₁,u₂,...,u_m) i-th element.

Vector component units 342 is for u minimum calculation unit 341 obtained₁～u_mComposition m dimensional vector U (u₁, u₂,...,u_m), as threshold vector.

Second Threshold vector determines that module 305, for calculating the m dimensional vector minimum so that L in equation below 4, is designated as U (u₁,u₂,...,u_m):

L=E{sign (PX+U)^Tsign(PX'+U)|R}-αE{sign(PX+U)^TSign (PX'+U) | Q} (formula 4)

Wherein, E{sign (PX+U)^TSign (PX'+U) | Q} represents that high dimensional feature vector similar in described Q becomes through P Change and after described threshold vector compares and determines sign symbol, the average of the distance between the symbolic vector obtained；E{sign (PX+U)^TSign (PX'+U) | R} represents that in described R, inhomogeneous high dimensional feature vector is through P conversion and through described threshold value After vector compares and determines sign symbol, the average of the distance between the symbolic vector obtained；

Second Threshold vector determines that module 305 is to U (u₁,u₂,...,u_m) be optimized after, obtain described threshold vector.

Above-mentioned Second Threshold vector determines that module 305 specifically includes: minimum calculation unit 351, vector optimization unit 352 and vector component units 353.

Minimum calculation unit 351 is identical with the function of above-mentioned minimum calculation unit 341, and here is omitted.

Vector optimization unit 352 is for the U (u asking for minimum calculation unit 351₁,u₂,...,u_m) element u_iValue It is optimized: for the element u of described threshold vector U_i, utilize equation below 5 and formula 6, ask for so that FN (u_i)+α×FP (u_i) minimum u_iValue, as the u after optimizing_iValue；

FN(u_i)=Pr (min{z, z'} >=u_i or max{z,z'}<u_i| Q) (formula 5)

FP(u_i)=Pr (min{z, z'} <u_i≤ max{z, z'} | R) (formula 6)

In described formula 5, (min{z, z'} >=u_i or max{z,z'}<u_i| Q) in z and z' represent in described Q arbitrarily A pair similar high dimensional feature vector X and X' in one set element obtains after converting respectively through described projection matrix P The i-th element of vector, Pr (min{z, z'} >=u_i or max{z,z'}<u_i| Q) represent for the set element in described Q, u_iMeet following condition: min{z, z'} >=u_i or max{z,z'}<u_iProbability；

In described formula 6, (min{z, z'} <u_i≤ max{z, z'} | R) in z and z' represent in described R any one collection Close the vector obtained after a pair inhomogeneous high dimensional feature vector X and X' in element converts respectively through described projection matrix P I-th element, Pr (min{z, z'} <u_i≤ max{z, z'} | R) represent for the set element in described R, u_iMeet as follows Condition: min{z, z'} <u_iThe probability of≤max{z, z'}.

The vector component units 353 u after vector optimization unit 352 is optimized₁～u_mForm described threshold vector.

As shown in Figure 4, can carry out according to existing multimedia information retrieval method based on the design of segmented index thought The retrieval of multimedia messages based on bit vectors, to obtain retrieving result, specifically includes following steps:

S401: extract the characteristic of present multimedia information, the high dimensional feature that the n of present multimedia information is tieed up to Amount is mapped as the bit vectors of m dimension, obtains the bit vectors of present multimedia information.

Specifically, after extracting the characteristic of present multimedia information, the method using the invention described above, will be the most The bit vectors that high dimensional feature DUAL PROBLEMS OF VECTOR MAPPING is m dimension of the n dimension of media information, obtains the bit vectors of present multimedia information.

S402: the bit vectors of present multimedia information is carried out even partition, obtains k son of present multimedia information Vector.

Specifically, the jth subvector of present multimedia information is by the bit vectors even partition of present multimedia information After jth group element composition, (the j-1) × v during wherein jth group element specifically includes the bit vectors of present multimedia information + 1 element～jth × l element；Wherein j is the natural number of 1～k, and v is the vector in each subvector (or every group element) Element number.

S403: for each subvector of present multimedia information, determines should the candidate collection of subvector respectively.

Specifically, for each subvector of present multimedia information, determine the candidate collection of correspondence respectively, so that it is determined that Go out k candidate collection；Wherein, during determining the candidate collection of jth subvector of corresponding present multimedia information, right In the jth subvector of described present multimedia information, the candidate collection of its correspondence determines according to following method: at jth rope The indexed set of guiding structure finds out the index identical with the jth subvector of this multimedia messages to be retrieved, and will find out The vectorial set corresponding to index as the candidate collection of the jth subvector of corresponding present multimedia information.

Wherein, bit vectors and the vectorial thereof of each multimedia messages to be retrieved is pre-stored within characteristics of the multimedia number According in storehouse, and for each multimedia messages to be retrieved, in advance the tag bit vector of this multimedia messages to be retrieved is carried out Even partition, sets up segmented index, obtains k index structure.

S404: for each vectorial in the candidate collection that obtains, find out in characteristics of the multimedia data base respectively Corresponding bit vectors.

Specifically, the Candidate Set of each subvector of corresponding present multimedia information for obtaining in above-mentioned steps S403 Close, i.e. k candidate collection, find out in characteristics of the multimedia data base correspondence candidate collection in each vectorial bit to Amount.

S405: calculate the Hamming distance between bit vectors and the bit vectors found of present multimedia information.

S406: Hamming distance is met the multimedia messages corresponding to the bit vectors imposed a condition defeated as retrieval result Go out.

Specifically, meet the bit vectors imposed a condition and specifically may is that bit vectors with present multimedia information The Hamming distance bit vectors less than or equal to d；More preferably, above-mentioned k is more than d, i.e. d less than or equal to k, and so can ensure that will not Missing inspection occur, the vectorial i.e. meeting the bit vectors imposed a condition is included in candidate collection.Generally, for satisfied retrieval Requirement, Hamming distance d value is set to a less number, the number of such as less than 3 or 4 by those skilled in the art；Therefore, generally V is at least double figures, the most greatly.

In sum, in technical solution of the present invention, owing to the high dimensional feature vector of present multimedia information is converted into bit Effect discrete between gathering, class in there is after vector class, thus ensure that original vector identification ability；So, application maturation The retrieval technique of bit vectors based on low-dimensional, it is possible to achieve compared to the higher inspection of retrieval technique based on high dimensional feature vector Rope efficiency, and less retrieval consumption, and make the retrieval result that the retrieval of multimedia messages based on bit vectors draws The most accurate, reduce the erroneous matching rate of retrieval.

The above is only the preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art For Yuan, under the premise without departing from the principles of the invention, it is also possible to make some improvements and modifications, these improvements and modifications also should It is considered as protection scope of the present invention.

Claims (13)

1. a multimedia information retrieval method based on bit vectors, it is characterised in that including:

After extracting the characteristic of present multimedia information, obtain the high dimensional feature vector of the n dimension of described present multimedia information, It is designated as X (x₁,x₂,...,x_n)；

By high dimensional feature vector X (x₁,x₂,...,x_n) by obtaining the intermediate vector W (w of m dimension after projection matrix P conversion₁, w₂,...,w_m)；

Each element respective element with described intermediate vector respectively of the threshold vector tieed up by m compares, according to comparative result Described intermediate vector is carried out binaryzation, obtains the bit vectors of the m dimension of described present multimedia information；Wherein, m is less than n；

According to the bit vectors obtained, characteristics of the multimedia data base finds out the bit vectors similar to this bit vectors, The multimedia messages corresponding to bit vectors that will find out is as retrieval result output；

Wherein, described projection matrix P is the matrix of m × n, and meets following condition: classify for each of storage in information bank The high dimensional feature vector of multimedia messages, the vectorial spacing expectation after P converts of the most similar high dimensional feature vector Value, minimum with the difference of the inhomogeneous high dimensional feature vector vectorial spacing expected value after P converts；

Described threshold vector meets following condition: for the high dimensional feature of each multimedia messages of storage in described information bank to Amount, the most similar high dimensional feature vector through P conversion and compare through described threshold vector, vectorial spacing after binaryzation From expected value, with inhomogeneous high dimensional feature vector through P conversion and compare through described threshold vector, after binaryzation to The difference of amount spacing expected value is minimum.

2. the method for claim 1, it is characterised in that described extraction present multimedia information characteristic it Before, also include:

Described projection matrix P is trained by the multimedia messages of storage in described information bank:

For the multimedia messages of storage in described information bank, by the high dimensional feature of wherein any pair similar multimedia messages Vector, as a set element, stores in similar sample set；And

Using the high dimensional feature vector of wherein any pair inhomogeneous multimedia messages as a set element, storage is to non-same In class sample set；

Construct so that in equation below 1Minimum projection matrix P:

Wherein, Q is described similar sample set；R is described non-similar sample set；E{||PX-PX'||²| Q} represents in described Q The similar high dimensional feature vector vectorial spacing expected value after P converts, X and X' therein represents any one collection in Q Close a pair similar high dimensional feature vector in element；E{||PX-PX'||²| R} represents inhomogeneous high dimensional feature in described R The vector vectorial spacing expected value after P converts, X and X' therein represents a pair in R in any one set element Inhomogeneous high dimensional feature vector；α is the weights set.

3. method as claimed in claim 2, it is characterised in that described in construct so that in equation below 1Minimum projection square Battle array P, specifically includes:

Ask for matrix Σ_GM minimum n tie up matrix characteristic vector；Wherein,Described Σ_QSuch as formula 2 institute Show, described Σ_RAs shown in Equation 3:

Σ_Q=E{ (X-X') (X-X')^T| Q} (formula 2)

In described formula 2, E{ (X-X') (X-X')^T| Q} represents the covariance matrix between high dimensional feature vector similar in described Q Average；

Σ_R=E{ (X-X') (X-X')^T| R} (formula 3)

In described formula 3, E{ (X-X') (X-X')^T| R} represents the covariance square in described R between inhomogeneous high dimensional feature vector The average of battle array；

Tieed up matrix characteristic vector by m the n asked for, constitute the projection matrix P of m × n.

4. method as claimed in claim 2, it is characterised in that described by described information bank in the multimedia messages of storage After training described projection matrix P, also include:

Calculate the m dimensional vector minimum so that L in equation below 4, be designated as U (u₁,u₂,...,u_m), and as described threshold value to Amount:

L=E{sign (PX+U)^Tsign(PX'+U)|R}-αE{sign(PX+U)^TSign (PX'+U) | Q} (formula 4)

Wherein, E{sign (PX+U)^TSign (PX'+U) | Q} represent similar in described Q high dimensional feature vector through P conversion and After described threshold vector compares and determines sign symbol, the average of the distance between the symbolic vector obtained；E{sign(PX+ U)^TSign (PX'+U) | R} represents that in described R, inhomogeneous high dimensional feature vector is through P conversion and through described threshold vector After relatively determining sign symbol, the average of the distance between the symbolic vector obtained.

5. method as claimed in claim 2, it is characterised in that described by described information bank in the multimedia messages of storage After training described projection matrix P, also include:

Calculate the m dimensional vector minimum so that L in equation below 4, be designated as U (u₁,u₂,...,u_m):

L=E{sign (PX+U)^Tsign(PX'+U)|R}-αE{sign(PX+U)^TSign (PX'+U) | Q} (formula 4)

Wherein, E{sign (PX+U)^TSign (PX'+U) | Q} represent similar in described Q high dimensional feature vector through P conversion and After described threshold vector compares and determines sign symbol, the average of the distance between the symbolic vector obtained；E{sign(PX+ U)^TSign (PX'+U) | R} represents that in described R, inhomogeneous high dimensional feature vector is through P conversion and through described threshold vector After relatively determining sign symbol, the average of the distance between the symbolic vector obtained；

Afterwards, to U (u₁,u₂,...,u_m) be optimized after, obtain described threshold vector:

Element u for described threshold vector U_i, utilize equation below 5 and formula 6, ask for so that FN (u_i)+α×FP(u_i) Little u_iValue, as the u after optimizing_iValue；

FN(u_i)=Pr (min{z, z'} >=u_iOr max{z, z'} ＜ u_i| Q) (formula 5)

FP(u_i)=Pr (min{z, z'} ＜ u_i≤ max{z, z'} | R) (formula 6)

In described formula 5, (min{z, z'} >=u_iOr max{z, z'} ＜ u_i| Q) in z and z' represent any one in described Q The vector that a pair similar high dimensional feature vector X and X' in set element obtains after converting respectively through described projection matrix P I-th element, Pr (min{z, z'} >=u_iOr max{z, z'} ＜ u_i| Q) represent for the set element in described Q, u_iFull The following condition of foot: min{z, z'} >=u_iOr max{z, z'} ＜ u_iProbability；

In described formula 6, (min{z, z'} ＜ u_i≤ max{z, z'} | R) in z and z' represent in described R any one set unit The i-th of the vector that a pair inhomogeneous high dimensional feature vector X and X' in element obtains after converting respectively through described projection matrix P Individual element, Pr (min{z, z'} ＜ u_i≤ max{z, z'} | R) represent for the set element in described R, u_iMeet following condition: Min{z, z'} ＜ u_iThe probability of≤max{z, z'}.

6. the method as described in claim 4 or 5, it is characterised in that described in calculate the m dimensional vector minimum so that L, specifically wrap Include:

Ask for so that the u of following expression 7 minimum_iValue；Wherein, i is the natural number of 1～m；

E{sign((P_i ^TX+u_i)(P_i ^TX'+u_i))|R}-αE{sign((P_i ^TX+u_i)^T(P_i ^TX'+u_i)) | Q} (expression formula 7)

Wherein, P_i ^TThe i-th row vector for described projection matrix P；u_iFor U (u₁,u₂,...,u_m) i-th element；

And the u that will obtain₁～u_mForm described m dimensional vector.

7. a Multimedia information retrieval system based on bit vectors, it is characterised in that including:

Bit vectors modular converter, after the characteristic extracting present multimedia information, obtains described present multimedia letter The high dimensional feature vector of the n dimension of breath, is designated as X (x₁,x₂,...,x_n)；By high dimensional feature vector X (x₁,x₂,...,x_n) by projection Intermediate vector W (the w of m dimension is obtained after matrix P conversion₁,w₂,...,w_mAfter), by m tie up threshold vector each element respectively with institute The respective element stating intermediate vector compares, and according to comparative result, described intermediate vector is carried out binaryzation, obtains described working as The bit vectors of the m dimension of front multimedia messages；Wherein, m is less than n；

Retrieval module, the bit vectors of the present multimedia information for obtaining according to described bit vectors modular converter, many Media characteristic data base finds out the bit vectors similar to this bit vectors, many corresponding to the bit vectors that will find out Media information is as retrieval result output；

Wherein, described projection matrix P is the matrix of m × n, and meets following condition: classify for each of storage in information bank The high dimensional feature vector of multimedia messages, the vectorial spacing expectation after P converts of the most similar high dimensional feature vector Value, minimum with the difference of the inhomogeneous high dimensional feature vector vectorial spacing expected value after P converts；

Described threshold vector meets following condition: for the high dimensional feature of each multimedia messages of storage in described information bank to Amount, the most similar high dimensional feature vector through P conversion and compare through described threshold vector, vectorial spacing after binaryzation From expected value, with inhomogeneous high dimensional feature vector through P conversion and compare through described threshold vector, after binaryzation to The difference of amount spacing expected value is minimum.

8. system as claimed in claim 7, it is characterised in that described bit vectors modular converter specifically includes:

High dimensional feature vector determination unit, after the characteristic extracting present multimedia information, obtains described current many matchmakers The high dimensional feature vector of the n dimension of body information, is designated as X (x₁,x₂,...,x_n)；

Intermediate vector computing unit, for the high dimensional feature vector X (x obtained by described high dimensional feature vector determination unit₁, x₂,...,x_n) by obtaining the intermediate vector W (w of m dimension after projection matrix P conversion₁,w₂,...,w_m)；

Threshold value comparing unit, each element of the threshold vector for being tieed up by m obtains with described intermediate vector computing unit respectively The respective element of intermediate vector compares, and according to comparative result, described intermediate vector is carried out binaryzation, obtain described currently The bit vectors of the m dimension of multimedia messages；Wherein, m is less than n.

9. system as claimed in claim 8, it is characterised in that also include:

Projection matrix builds module, for training described projection matrix P by the multimedia messages of storage in described information bank: For the multimedia messages of storage in described information bank, by the high dimensional feature vector of wherein any pair similar multimedia messages As a set element, store in similar sample set；And by the height of wherein any pair inhomogeneous multimedia messages Dimensional feature vector, as a set element, stores in non-similar sample set；Construct so that in equation below 1Minimum Projection matrix P:

Wherein, Q is described similar sample set；R is described non-similar sample set；E{||PX-PX'||²| Q} represents in described Q The similar high dimensional feature vector vectorial spacing expected value after P converts, X and X' therein represents any one collection in Q Close a pair similar high dimensional feature vector in element；E{||PX-PX'||²| R} represents inhomogeneous high dimensional feature in described R The vector vectorial spacing expected value after P converts, X and X' therein represents a pair in R in any one set element Inhomogeneous high dimensional feature vector；α is the weights set；

First threshold vector determines module, for calculating the m dimensional vector minimum so that L in equation below 4, is designated as U (u₁, u₂,...,u_m), and as described threshold vector:

L=E{sign (PX+U)^Tsign(PX'+U)|R}-αE{sign(PX+U)^TSign (PX'+U) | Q} (formula 4)

Wherein, E{sign (PX+U)^TSign (PX'+U) | Q} represent similar in described Q high dimensional feature vector through P conversion and After described threshold vector compares and determines sign symbol, the average of the distance between the symbolic vector obtained；E{sign(PX+ U)^TSign (PX'+U) | R} represents that in described R, inhomogeneous high dimensional feature vector is through P conversion and through described threshold vector After relatively determining sign symbol, the average of the distance between the symbolic vector obtained.

10. system as claimed in claim 9, it is characterised in that described first threshold vector determines that module specifically includes:

Minimum calculation unit, for asking for so that the u of following expression 7 minimum_iValue；Wherein, i is the natural number of 1～m；

E{sign((P_i ^TX+u_i)(P_i ^TX'+u_i))|R}-αE{sign((P_i ^TX+u_i)^T(P_i ^TX'+u_i)) | Q} (expression formula 7)

Wherein, P_i ^TThe i-th row vector for described projection matrix P；u_iFor U (u₁,u₂,...,u_m) i-th element；

Vector component units, for the u described minimum calculation unit obtained₁～u_mForm described m dimensional vector U (u₁, u₂,...,u_m), as described threshold vector.

11. systems as claimed in claim 8, it is characterised in that also include:

Projection matrix builds module, for training described projection matrix P by the multimedia messages of storage in described information bank: For the multimedia messages of storage in described information bank, by the high dimensional feature vector of wherein any pair similar multimedia messages As a set element, store in similar sample set；And by the height of wherein any pair inhomogeneous multimedia messages Dimensional feature vector, as a set element, stores in non-similar sample set；Construct so that in equation below 1Minimum Projection matrix P:

Wherein, Q is described similar sample set；R is described non-similar sample set；E{||PX-PX'||²| Q} represents in described Q The similar high dimensional feature vector vectorial spacing expected value after P converts, X and X' therein represents any one collection in Q Close a pair similar high dimensional feature vector in element；E{||PX-PX'||²| R} represents inhomogeneous high dimensional feature in described R The vector vectorial spacing expected value after P converts, X and X' therein represents a pair in R in any one set element Inhomogeneous high dimensional feature vector；α is the weights set；

Second Threshold vector determines module, for calculating the m dimensional vector minimum so that L in equation below 4, is designated as U (u₁, u₂,...,u_m):

L=E{sign (PX+U)^Tsign(PX'+U)|R}-αE{sign(PX+U)^TSign (PX'+U) | Q} (formula 4)

Wherein, E{sign (PX+U)^TSign (PX'+U) | Q} represent similar in described Q high dimensional feature vector through P conversion and After described threshold vector compares and determines sign symbol, the average of the distance between the symbolic vector obtained；E{sign(PX+ U)^TSign (PX'+U) | R} represents that in described R, inhomogeneous high dimensional feature vector is through P conversion and through described threshold vector After relatively determining sign symbol, the average of the distance between the symbolic vector obtained；

Second Threshold vector determines that module is to U (u₁,u₂,...,u_m) be optimized after, obtain described threshold vector.

12. systems as claimed in claim 11, it is characterised in that described Second Threshold vector determines that module specifically includes:

Minimum calculation unit, for asking for so that the u of following expression 7 minimum_iValue；Wherein, i is the natural number of 1～m；

E{sign((P_i ^TX+u_i)(P_i ^TX'+u_i))|R}-αE{sign((P_i ^TX+u_i)^T(P_i ^TX'+u_i)) | Q} (expression formula 7)

Wherein, P_i ^TThe i-th row vector for described projection matrix P；u_iFor U (u₁,u₂,...,u_m) i-th element；

Vector optimization unit, for U (u₁,u₂,...,u_m) element u_iIt is optimized: for the element of described threshold vector U u_i, utilize equation below 5 and formula 6, ask for so that FN (u_i)+α×FP(u_i) minimum u_iValue, as the u after optimizing_iValue；

FN(u_i)=Pr (min{z, z'} >=u_iOr max{z, z'} ＜ u_i| Q) (formula 5)

FP(u_i)=Pr (min{z, z'} ＜ u_i≤ max{z, z'} | R) (formula 6)

In described formula 5, (min{z, z'} >=u_iOr max{z, z'} ＜ u_i| Q) in z and z' represent any one in described Q The vector that a pair similar high dimensional feature vector X and X' in set element obtains after converting respectively through described projection matrix P I-th element, Pr (min{z, z'} >=u_iOr max{z, z'} ＜ u_i| Q) represent for the set element in described Q, u_iFull The following condition of foot: min{z, z'} >=u_iOr max{z, z'} ＜ u_iProbability；

In described formula 6, (min{z, z'} ＜ u_i≤ max{z, z'} | R) in z and z' represent in described R any one set unit The i-th of the vector that a pair inhomogeneous high dimensional feature vector X and X' in element obtains after converting respectively through described projection matrix P Individual element, Pr (min{z, z'} ＜ u_i≤ max{z, z'} | R) represent for the set element in described R, u_iMeet following condition: Min{z, z'} ＜ u_iThe probability of≤max{z, z'}；

Vector component units, for by the u after described vector optimization unit optimization₁～u_mForm described threshold vector.

13. systems as described in claim 7-12 is arbitrary, it is characterised in that projection matrix builds module and specifically includes:

Minimum matrix characteristic vector computing unit, is used for asking for matrix Σ_GM minimum n tie up matrix characteristic vector；Wherein,Described Σ_QAs shown in Equation 2, described Σ_RAs shown in Equation 3:

Σ_Q=E{ (X-X') (X-X')^T| Q} (formula 2)

In described formula 2, E{ (X-X') (X-X')^T| Q} represents the covariance matrix between high dimensional feature vector similar in described Q Average, a pair similar high dimensional feature vector that X and X' therein represents in Q in any one set element；

∑_R=E{ (X-X') (X-X')^T| R} (formula 3)

In described formula 3, E{ (X-X') (X-X')^T| R} represents the covariance square in described R between inhomogeneous high dimensional feature vector The average of battle array, a pair inhomogeneous high dimensional feature vector that X and X' therein represents in R in any one set element；

Projection matrix determines unit, for being tieed up matrix characteristic vector by m the n asked for, constitutes the projection matrix P of m × n.