CN102426606A - Method for retrieving multi-feature image based on particle swarm algorithm - Google Patents

Abstract

The invention discloses a method for retrieving a multi-feature image based on a particle swarm algorithm. The method comprises the following steps of: firstly, according to a retrieved image, extracting a plurality of feature vectors of the retrieved image, and normalizing, thereby acquiring a feature distance; fusing the feature distance; acquiring a comprehensive feature fused distance according to the optimized weight distribution; retrieving according to image similarity; feeding a result back to a user; and retrieving the image similar to the retrieved image. According to image information provided by the user, the multi-feature of the image is extracted and a multi-feature distance is calculated; the particle swarm algorithm is utilized to optimize parameters; the multi-feature distance is fused and then is matched with an appointed image feature bank; the image similar to the retrieved image is efficiently retrieved; the retrieval accuracy is promoted; and the method is conveniently adopted by the user.

Description

A kind of multi-feature image retrieval method based on particle cluster algorithm

Technical field

The invention belongs to digital image processing field, relate to a kind of multi-feature image retrieval method based on particle cluster algorithm.

Background technology

CBIR (content-based image retrieval; CBIR) more and more become a popular research field both domestic and external; It has merged the technological achievement in fields such as Flame Image Process, image recognition and image data base; Make full use of the index that low-level image features such as self-contained characteristic attribute of image such as color, texture, shape set up and retrieve, thereby image retrieval means more effectively more accurately can be provided.The single image characteristic can provide the part authentication information of image, but insufficient, and many characteristics of image are carried out Combination application authentication information fully can be provided relatively comprehensively.Based on to the taking all factors into consideration of local characteristics of image and many characteristics combination, the image retrieval algorithm of many characteristics combination has outstanding retrieval performance, and its result for retrieval more meets the visual experience of human eye.And the weight of how to distribute the different images characteristic is a problem.

At present, image retrieval appraisal procedure commonly used is the PVR curve, and the PVR curve is by recall ratio and precision ratio definition.Suppose that A is the image collection relevant with query image all in the test pattern storehouse, the image collection that B returns for retrieval.A is the number of the associated picture that inquires; B is the number of the uncorrelated image that inquires; C is relevant with detected image, but the number that does not retrieve.

Being defined as of recall ratio:

$\mathrm{recall}=P\left(B\right|A)=\frac{a}{a+c};$

(1)

Being defined as of precision ratio:

$\mathrm{precision}=\frac{P(A\∩B)}{P\left(B\right)}=\frac{a}{a+b};$

(2)

Behind the recall ratio and precision ratio of the result for retrieval that calculates piece image, if with the x axle of recall ratio as coordinate, precision ratio then can be drawn out the PVR curve of image searching result as the y axle of coordinate.If the PVR curve is f (x; Y); F (x then; Y) area that surrounds with the x-y axle claims that for

S (f) is the PVR index, is designated as E.Then E is big more, and image retrieval performance is good more; E is more little, and image retrieval performance is poor more.If E=1, the image retrieval performance reaches best so, its PVR curve be f (x, y)=1.

Particle swarm optimization algorithm PSO (particle swarm optimization) is a kind of algorithm based on the intelligence of trooping, and it through the information sharing between the single particle, seeks the optimum point in complex search space through the foraging behavior of the simulation flock of birds or the shoal of fish.The PSO algorithm is a kind of optimisation technique based on the population operation.As far as optimization problem, each particle is represented possible separating in the PSO algorithm.Each particle is in the desired positions that iterative process lived through in the colony, is exactly preferably separating of being found of this particle itself.The desired positions that whole colony lived through, what just the whole colony of foot found at present preferably separates.The former is called individual extreme value, and the latter is called global extremum.Each particle is all brought in constant renewal in oneself through above-mentioned two extreme values, thereby produces colony of new generation, and just whole colony carries out thorough search to separating the zone in this process.

If the population size of particle is n, then i (i=1,2 ..., n) individual particle position can be expressed as x _i, its individual extreme value is designated as pBest _i, its speed is used v _iExpression, the global extremum of colony is represented with gBest.So arbitrary particle i will upgrade oneself speed and position according to following formula:

v _i(t+1)＝ωv _i(t)+c ₁r ₁(t)(pBest _i(t)-x _i(t))+c ₂r ₂(t)(gBest _i(t)-x _i(t))

(3)

x _i(t+1)＝x _i(t)+v _i(t+1)

(4)

C wherein ₁, c ₂Be constant, be called the study factor or accelerator coefficient; r ₁And r ₂It is the random number on (0,1); ω is inertia weight (inertia weight).

The individual extreme value of each particle is upgraded with following formula:

${\mathrm{pBest}}_i(t+1)=\left\{\begin{array}{cc}{x}_i(t+1),\mathrm{if}& x_i(t+1)\&GreaterEqual;{\mathrm{pBest}}_i\left(t\right)\\ \mathrm{pBes}{t}_i\left(t\right),\mathrm{if}& x_i(t+1)<{\mathrm{pBest}}_i\left(t\right)\end{array}\right.$

(5)

Global extremum to all particles is chosen as follows:

gBest(t+1)＝max(pBest _i(t+1))，(i＝1，2，…，n)

(6)

Value to each particle's velocity q is limited at [v _Max, v _Max] in, v _MaxValue get the width of search volume usually. the setting that inertia weight is added normally is reduced to 0.2 from 0.9 linearity.

Formula (3) is made up of three parts, and first is the previous speed of particle, and the state that particle is present has been described; Second portion is a cognitive part (Cognition Modal).The thinking of expression particle itself; Third part is society's part (Social Modal).Three parts have determined the space search ability of particle jointly.The ability of the balance overall situation and Local Search has played in first.Second portion makes particle that enough strong ability of searching optimum arranged, and avoids local minimum.Third part has embodied interparticle information sharing.

Summary of the invention

Weights optimum problem how the object of the present invention is to provide a kind of multi-feature image retrieval method based on particle cluster algorithm when solving that many characteristics of image merge in the image retrieval.The image information that this method can provide according to the user goes out many characteristic distances through many feature calculation of extracting image, utilizes particle cluster algorithm to optimize weighting parameter, and many characteristic distances are merged, and matees with the characteristics of image storehouse of appointment then.Thereby retrieve efficiently and the image of the image similarity that is retrieved, improve retrieval rate, satisfy customer requirements.

The objective of the invention is to realize through following technical scheme:

A kind of multi-feature image retrieval method based on particle cluster algorithm; It is characterized in that: the image information that this method provides according to the user; Many feature calculation through extracting image go out many characteristic distances, utilize particle cluster algorithm to optimize weighting parameter, and many characteristic distances are merged; Mate with the characteristics of image storehouse of appointment then, thereby retrieve efficiently and the image of the image similarity that is retrieved; Specifically comprise the steps:

Step 1: establish P _tBe the image that to retrieve, N width of cloth image arranged, P in the image data base _iBe i width of cloth image in the shape library, 1≤i≤N;

Step 2: extract a plurality of proper vectors of image that are retrieved, normalization, characteristic set is W={W ₁, W ₂..., W _K, wherein K is the characteristic number of extracting, K>=1;

Step 3: obtain P according to following formula _t, P _iAbout characteristic W _KCharacteristic distance D _i

$D(q,t)={\left(\underset{m=0}{\overset{M-1}{\mathrm{\&Sigma;}}}{|w_q\left(m\right)-w_t\left(m\right)|}^2\right)}^{\frac{1}{2}}$

(7)

Step 4: with distance B i normalization, obtain D according to following formula _i' mistake! Do not find Reference source.；

$d_i^{\&prime;}=\left\{\begin{array}{cc}\frac{d_i-\mathrm{mD}+3\mathrm{\&sigma;}}{6\mathrm{\&sigma;}}& d_i^{\&prime;}\&Element;(\mathrm{mD}-3\mathrm{\&sigma;},\mathrm{mD}+3\mathrm{\&sigma;})\\ 0& d_i^{\&prime;}\&le;\mathrm{mD}-3\mathrm{\&sigma;}\\ 1& d_i^{\&prime;}\&GreaterEqual;\mathrm{mD}-3\mathrm{\&sigma;}\end{array}\right.$

(8)

Wherein $\mathrm{MD}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{d}_i|n=\mathrm{1,2},...,N$

${\mathrm{\&sigma;}}^2=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{(d_i-\mathrm{mD})}^{2}n=\mathrm{1,2},...,N$

Step 5: K characteristic distance merged according to following formula;

D _Is=x ₁D ' ₁+ x ₂D ' ₂+ ... + x _KD ' _K, x wherein ₁..., x _K∈ [0,1], and x ₁+ ... + x _K=1; (9)

Utilize particle cluster algorithm to carry out parameter optimization, obtain optimized weights and distribute Xbest;

Step 6: after distributing Xbest according to optimized weights, obtain the comprehensive characteristics fusion distance, and retrieve, the result is fed back to the user, retrieve and the image of the image similarity that is retrieved according to image similarity.

Further providing below utilizes particle cluster algorithm to carry out the step of parameter optimization:

Step1: suppose that search space is that K ties up and population has m particle, i particle represented the parameter vector X of a K dimension _i=(x _I1, x _I2..., x _IK), (i=1,2 ..., m), promptly i particle is X in the position of the search volume of K dimension _iIn other words, each particle position is the feasible solution that potential weights distribute, a random initializtion in allowed limits.

Step2: with X _iObjective function of substitution just can calculate its fitness, the quality of weighing according to the size of fitness.Objective function f (X _i) be made as many signature searchs result's under the current weight PVR index.

Step3: i the particle speed of " circling in the air " also is the vector of a K dimension, is expressed as: V _i=(v _I1, v _I2..., v _IK), (i=1,2 ..., m), random initializtion in allowed limits.

Step4: the optimal location note that i particle oneself searches is made P _i=(p _I1, p _I2..., p _IK), (i=1,2 ..., m).The P of each particle _iThe coordinate initialization is set to its current location, and calculates the fitness value (weights that are current individual representative distribute the PVR index result who retrieves) of its corresponding individual extreme point.

Step5: the optimal location note that whole population searches is up to now made P _g=(p _G1, p _G2..., p _GK).P _gBe initialized as in the step best individual body position in all individual extreme values, be current best weight value distribution.

Step6: will come particle is carried out iterative operation according to following formula:

V’ _ik＝ω·V _ik+c ₁·rand ₁·(P _ik-X _ik)+c ₂·rand ₂·(P _gk-X _ik)

(10)

X’ _ik＝X _ik+V _ik

(11)

I=1 wherein, 2 ..., m, ω is an inertia weight, is a constant between [0,1]; c ₁And c ₂Being learning rate, also is a non-negative constant; Rand ₁And rand ₂It is the random number that produces between [0,1]; V _Ik∈ [V _Max, V _Max], and V _MaxIt is the speed maximal value of an appointment.Can find out that from top two formula the moving direction of particle is by three part decisions, own original speed V _Ik, with the range difference (P of the own optimum position that experiences _Ik-X _Ik) and with the range difference (P of the optimum position of colony experience _Gk-X _Ik), and respectively by weight coefficient ω, c ₁And c ₂Determine its relative importance.The standard that iteration stops is the optimal-adaptive degree that perhaps reaches appointment according to maximum iteration time.

Step7: the individual extreme value of each particle is upgraded with following formula:

${P_i}^{\&prime;}=\left\{\begin{array}{cc}{X_i}^{\&prime;},\mathrm{if}& f\left({X_i}^{\&prime;}\right)\&GreaterEqual;f\left(P_i\right)\\ P_i,\mathrm{if}& f\left({X_i}^{\&prime;}\right)<f\left(P_i\right)\end{array}\right.$

(12)

Estimate each particle, calculate the fitness value of particle,, then upgrade this particle position if be better than the current individual extreme value of this particle.

Step8: the global extremum to all particles is chosen as follows:

P _g′＝P _imax′

(13)

P wherein _Imax' be the maximum Pi ' of adaptive value.

Step9: new particle more, whether check meets termination condition, if current iterations has reached predefined maximum times (or reach least error require), then stops iteration, the output optimum solution, otherwise continue more new particle.

The present invention has overcome the optimization problem of weighting parameter when many characteristics of image merge in the image retrieval; The image information that can provide according to the user; Many feature calculation through extracting image go out many characteristic distances; Utilize particle cluster algorithm to optimize weighting parameter, many characteristic distances are merged, mate with the characteristics of image storehouse of appointment then.Thereby retrieve efficiently and the image of the image similarity that is retrieved, improve retrieval rate, satisfy customer requirements.

The present invention is applicable in the Digital Image Processing, can retrieve efficiently and the image of the image similarity that is retrieved, and retrieval rate is high.

Description of drawings

Fig. 1 carries out the schematic flow sheet of image retrieval for the present invention.

Fig. 2 is 100 images randomly drawing in 1000 trademark image storehouses.

Fig. 3, Fig. 5, Fig. 7 are 3 sub-class libraries in trademark image storehouse.

Fig. 4, Fig. 6, Fig. 8 are single characteristic key PVR curve of averaging of income and average comprehensive characteristics retrieval PVR curve behind the 3 sub-category library searchings.

HU, ZERNIKE, LEGENDRE represent the result for retrieval based on Hu square, zernike square and legendre square respectively among the figure; The GFD representative is based on the result for retrieval of broad sense fourier descriptors; The ENTROPY representative is comprehensively represented based on the result for retrieval after many characteristic distances fusions of particle cluster algorithm based on information entropy characteristic key result.

Embodiment

Below in conjunction with accompanying drawing the present invention is elaborated, but protection scope of the present invention is not limited only to this.

Embodiment 1

A kind of image search method of comprehensive a plurality of region shape characteristics; The selected characteristic descriptor comprises: Hu square, Zernike square, Legendre square, generalized Fourier descriptor (generic Fourier descriptor; GFD) and the information entropy characteristic, introduce as follows respectively:

1.1Hu square

Hu has proposed invariant moments first in 1962 theoretical, is characterized in being satisfied with image translation, flexible and invariable rotary, and square is used for shape recognition.That Hu utilizes is normalized two, third central moment has been constructed 7 invariant moments that translation, rotation and change of scale had unchangeability.Invariant moments is a kind of statistical nature of image, and as in probability, replacing its distribution law to describe the statistical nature of this stochastic variable with each rank square of stochastic variable, each rank square that it utilizes gradation of image to distribute is described the characteristic that gradation of image distributes.

Being defined as of geometric invariant moment: suppose a bianry image be f (x, y), width is W, highly is H, its p+q rank square is:

$m_{\mathrm{pq}}=\underset{x=0}{\overset{W-1}{\mathrm{\&Sigma;}}}\underset{y=0}{\overset{H-1}{\mathrm{\&Sigma;}}}{x}^p{y}^q*f(x,y),p,q=\mathrm{0,1,2},...$ (14)

Its centre distance is defined as:

${\mathrm{\&mu;}}_{\mathrm{pq}}=\frac{1}{m_{00}^{\frac{p+q}{2}+1}}\underset{x=0}{\overset{W-1}{\mathrm{\&Sigma;}}}\underset{y=0}{\overset{H-1}{\mathrm{\&Sigma;}}}{(x-x_0)}^p{(y-y_0)}^q*f(x,y),p,q=\mathrm{0,1,2},...$ (15)

In the formula $x_0=\frac{m_{10}}{m_{00}},$ ${\mathrm{<figure-callout\; id="0"\; label="y"\; filenames="111111163403.png"\; state="\{\{state\}\}">y</figure-callout>}}_0=\frac{m_{01}}{m_{00}},$ (x ₀, y ₀) be the barycenter of regional graphics.

μ _PqUnchangeability with Pan and Zoom, but,, eliminate rotation difference therefore through combination second-order moment around mean and third central moment to rotating sensitivity, obtain 7 invariant features squares:

η ₁＝μ ₀₂+μ ₂₀

(16)

${\mathrm{\&eta;}}_2={({\mathrm{\&mu;}}_{20}-{\mathrm{\&mu;}}_{02})}^2+{4\mathrm{\&mu;}}_{11}^2$

(17)

η ₃＝(μ ₃₀-3μ ₁₂) ²+(3μ ₂₁-μ ₀₃) ²

(18)

η ₄＝(μ ₃₀+μ ₁₂) ²+(μ ₂₁+μ ₀₃) ²

(19)

η ₅＝(μ ₃₀-3μ ₁₂)(μ ₃₀+μ ₁₂)[(μ ₃₀+μ ₁₂) ²-3(μ ₂₁+μ ₀₃) ²]+(3μ ₂₁-μ ₀₃)(μ ₂₁+μ ₀₃)[3(μ ₃₀+μ ₁₂) ²-(μ ₂₁+μ ₀₃) ²]

(20)

η ₆＝(μ ₂₀-μ ₀₂)[(μ ₃₀+μ ₁₂) ²-(μ ₂₁+μ ₀₃) ²]+4μ ₁₁(μ ₃₀+μ ₁₂)(μ ₂₁+μ ₀₃)

(21)

η ₇＝(3μ ₂₁-μ ₀₃)(μ ₃₀+μ ₁₂)[(μ ₃₀+μ ₁₂) ²-3(μ ₂₁+μ ₀₃) ²]+(3μ ₁₂-μ ₃₀)(μ ₂₁+μ ₀₃)[3(μ ₃₀+μ ₁₂) ²-(μ ₂₁+μ ₀₃) ²]

(22)

Therefore the characteristics of image vector with unchangeability that is produced by geometric invariant moment is only 7; The advantage of using geometric invariant moment be it be a very compact shape represent and also calculated amount little; Shortcoming is that the minority invariant that has only the low order square to derive is not enough to describe accurately shape, and the invariant moments of high-order is difficult to again obtain.

1.2Zernike square

1980, Irishman Teague was the basis with the Zernike orthogonal polynomial, had provided two-dimensional function f (x, the definition of Zernike matrix y).Plural number Zernike square obtains from the Zernike polynomial expression:

$V_{\mathrm{pq}}(x,y)=V_{\mathrm{pq}}(r\cos \mathrm{\&theta;},r\sin \mathrm{\&theta;})=R_{\mathrm{pq}}\left(r\right)\mathrm{edp}\left(\hat{j}\mathrm{q\&theta;}\right)$ (23)

In the formula, r is that (x is y) to the radius of shape barycenter for point; θ is the angle of r and x axle; P, q are integer, and 0≤| q|≤p, p-|q| are even number.The Zernike polynomial expression is the complete quadrature plural number of a group in a unit circle base.(p, q) being defined as of rank plural number Zernike square:

$Z_{\mathrm{pq}}=\frac{p-1}{\mathrm{\&pi;}}\underset{x}{\mathrm{\&Sigma;}}\underset{y}{\mathrm{\&Sigma;}}f(x,y)V_{\mathrm{pq}}^{*}(x,y),(x^2+y^2\&le;1)$ (24)

(x y) is bianry image to f in the formula.Because the field of definition of Zernike basis function in unit circle, so calculate before the Zernike square, needs to specify unit circle, to obtain the unchangeability of yardstick and translation, utilizes the amplitude of Zernike square can arrive rotational invariance again.

For the image of H * W, the definition of Zernike square is following:

$Z_{\mathrm{pq}}={\mathrm{\&lambda;}}_z(p,N)\underset{i=0}{\overset{W-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{H-1}{\mathrm{\&Sigma;}}}{R}_{\mathrm{pq}}\left(r_{\mathrm{ij}}\right)\exp (-\hat{j}{\mathrm{q\&theta;}}_{\mathrm{ij}})f(i,j)$ (25)

Wherein $R_{\mathrm{Pq}}\left(r\right)=\underset{k=0}{\overset{(p-|q\left|\right)/2}{\mathrm{\&Sigma;}}}{(-1)}^k\frac{(p-k)!}{k!((p+|q\left|\right)/2-k)!((p-|q\left|\right)/2-k)!}{r}^{p-2k}$ (26)

With bianry image f (x, y) in the unit of the being mapped to garden, need do like down conversion:

$r_{\mathrm{ij}}=\sqrt{x_i^2+y_j^2},$ ${\mathrm{\&theta;}}_{\mathrm{ij}}={\tan }^{-1}\left(\frac{y_j}{x_i}\right),$ x _i＝c _1xi+c ₂，y _j＝c _1yj+c ₂ (27)

${\mathrm{\&lambda;}}_z(p,N)=\frac{P+1}{(W-1)(H-1)},$ $c_{1x}=\frac{2}{W-1},$ $c_{1y}=\frac{2}{H-1},$ c ₂＝-1 (28)

The Zernike square is a kind of plural square based on orthogonal polynomial, than other square littler information redundancy property is arranged, and has good rotational invariance, can construct any High Order Moment, with preceding 7 invariant composition characteristics vector of Zernike square.

1.3Legendre square

It is the Legendre square of kernel function that Teague has also proposed with the Legendre polynomial expression simultaneously, and the Legendre polynomial expression has constituted a complete orthogonal set in unit circle.(p+q) the Legendre square on rank is defined as:

$L_{\mathrm{pq}}=\frac{(2p+1)(2q+1)}{4}\underset{x}{\mathrm{\&Sigma;}}\underset{y}{\mathrm{\&Sigma;}}{p}_p\left(y\right)f(x,y)$ (29)

Wherein

is called the Legendre polynomial expression.Legendre is a kind of orthogonal moment, can describe the characteristic of image well, has reconfigurability, advantage that amount of redundant information is few, chooses 18 invariant composition characteristics vectors.

1.4 generalized Fourier descriptor

In order to overcome the shortcoming of Zernike square, Zhang and Lu proposed the generalized Fourier descriptor (genericFourier descriptor, GFD).GFD has adopted the plane polar coordinates Fourier transform of revising; Promptly image is carried out the polar coordinates sampling, the information of sampling is repainted under rectangular cartesian coordinate, again the image under this rectangular coordinate is done Fourier transform; Got an angular frequency territory, 4 amplitude frequency domain * 9, totally 36 frequency coefficients.

$\mathrm{GFD}=\{\frac{\left|\mathrm{PF}\left(\mathrm{0,0}\right)\right|}{\mathrm{area}},\frac{\left|\mathrm{PF}\left(\mathrm{0,1}\right)\right|}{\left|\mathrm{PF}\left(\mathrm{0,0}\right)\right|},...,\frac{\left|\mathrm{PF}(0,n)\right|}{\left|\mathrm{PF}\left(\mathrm{0,0}\right)\right|},...,\frac{\left|\mathrm{PF}(m,0)\right|}{\mathrm{PF}\left(\mathrm{0,0}\right)|},...,\frac{\left|\mathrm{PF}(m,n)\right|}{\left|\mathrm{PF}\left(\mathrm{0,0}\right)\right|}\}$ (30)

The GFD method has been chosen the amplitude information of these frequency coefficients, and does normalization and handle, and obtains having translation, the eigenmatrix of convergent-divergent and rotational invariance.

1.5 information entropy characteristic

One of founder of information theory Claude E.Shannon is defined as information entropy the probability of occurrence of Discrete Stochastic incident.So-called information entropy is abstract concept on the mathematics, can be understood as information entropy the probability of occurrence of certain customizing messages.In concrete computing information entropy, Shanon is defined as the ambiguity size that is eliminated in the system with its value, and ambiguity is described with random occurrence in theory of probability.For digital picture, the pixel of different gray scales is filled different zones with different probability distribution, thereby makes pictures different show different shape facilities.And for the two-value trademark image, its pixel grey scale has only 0 and 1 two kind of value, and therefore, its information entropy can be write as:

H(p ₀，p ₁)＝-p ₀logp ₀-p ₁logp ₁ (31)

Wherein, p ₀, p ₁It is respectively 0,1 two kind of probability that pixel occurs in image.

When the information entropy of computed image, for effectively reflection image distributed intelligence spatially, need image be carried out piecemeal, calculate the entropy of each subimage block then respectively.Piecemeal information entropy matrix description picture shape characteristic, its accuracy is relevant with the fine granularity of image block.That is to say that each subimage block is more little, the original image piecemeal is many more, and the description of characteristics of image is just accurate more.This paper carries out 64 * 64 piecemeal to 256 * 256 trademark image, and promptly each sub-block size is 4 * 4.Calculate the characteristic that the singular value of entropy matrix is come descriptor entropy matrix then.

1, choose the two-value trademark image of 1000 256x256 from MPEG, set up the trademark image database, Fig. 2 is 100 trademark images selecting at random.The trademark image storehouse is divided into 7 sub-class libraries.Image search method has the following steps, and sees Fig. 1:

Step 1: establish P _tBe the image that to retrieve, 1000 width of cloth images arranged, P in the image data base _iBe i width of cloth image in the shape library, 1≤i≤N.

Step 2: extract five proper vectors of the image that is retrieved, normalization, characteristic set is W={W ₁, W ₂..., W ₅.

Step 3: utilize following formula to calculate P _t, P _iAbout characteristic W _iCharacteristic distance D _i

$D(q,t)={\left(\underset{m=0}{\overset{M-1}{\mathrm{\&Sigma;}}}{|w_q\left(m\right)-w_t\left(m\right)|}^2\right)}^{\frac{1}{2}}$ (32)

Step 4: utilize following formula with distance B i normalization, obtain D _i' mistake! Do not find Reference source.。

$d_i^{\&prime;}=\left\{\begin{array}{cc}\frac{d_i-\mathrm{mD}+3\mathrm{\&sigma;}}{6\mathrm{\&sigma;}}& d_i^{\&prime;}\&Element;(\mathrm{mD}-3\mathrm{\&sigma;},\mathrm{mD}+3\mathrm{\&sigma;})\\ 0& d_i^{\&prime;}\&le;\mathrm{mD}-3\mathrm{\&sigma;}\\ 1& d_i^{\&prime;}\&GreaterEqual;\mathrm{mD}-3\mathrm{\&sigma;}\end{array}\right.$ (33)

Wherein $\mathrm{MD}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{d}_i|n=\mathrm{1,2},...,N$

${\mathrm{\&sigma;}}^2=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{(d_i-\mathrm{mD})}^{2}n=\mathrm{1,2},...,N$

Step 5: utilize following formula that 5 characteristic distances are merged.

D _Is=x ₁D ' ₁+ x ₂D ' ₂+ ... + x ₅D ' ₅, x wherein ₁..., x ₅∈ [0,1], and x ₁+ ... + x _K=1.(34)

Utilize particle cluster algorithm to carry out parameter optimization, obtain optimized weights and distribute Xbest.

Step 6: after obtaining optimized weights distribution Xbest, calculate the comprehensive characteristics fusion distance, retrieve according to image similarity again, calculate precision ratio and recall ratio, the result is fed back to the user.Present embodiment is chosen wherein Fig. 3, Fig. 5 and single characteristic key PVR curve of 3 sub-class libraries averagings of income and average comprehensive characteristics retrieval PVR curve shown in Figure 7, like Fig. 4, Fig. 6 and shown in Figure 8.This shows that single characteristic key has nothing in common with each other for the performance that different subclass storehouses is showed, and generally be superior to single characteristic key based on the retrieval performance of many characteristic key.

Further providing below utilizes particle cluster algorithm to carry out the step of parameter optimization:

Step1: suppose search space be 5 dimensions and establish population 30 particles arranged, i particle represented one 5 parameter vector X that ties up _i=(x _I1, x _I2..., x _I5), (i=1,2 ..., 30), promptly i particle is X in the position of the search volume of 5 dimensions _iIn other words, each particle position is the feasible solution that potential weights distribute, a random initializtion in allowed limits.

Step2: with X _iObjective function of substitution just can calculate its fitness, the quality of weighing according to the size of fitness.Objective function f (X _i) be made as many signature searchs result's under the current weight PVR index.

Step3: i the particle speed of " circling in the air " also is the vector of one 5 dimension, is expressed as: V _i=(v _I1, v _I2..., v _I5), (i=1,2 ..., 30), random initializtion in allowed limits.

Step4: the optimal location note that i particle oneself searches is made P _i=(p _I1, p _I2..., p _I5), (i=1,2 ..., 30).The P of each particle _iThe coordinate initialization is set to its current location, and calculates the fitness value (weights that are current individual representative distribute the PVR index result who retrieves) of its corresponding individual extreme point.

Step5: the optimal location note that whole population searches is up to now made P _g=(p _G1, p _G2..., p _G5).P _gBe initialized as in the step best individual body position in all individual extreme values, be current best weight value distribution.

Step6: will come particle is carried out iterative operation according to following formula:

V’ _ik＝ω·V _ik+c ₁·rand ₁·(P _ik-X _ik)+c ₂·rand ₂·(P _gk-X _ik) (35)

X’ _ik＝X _ik+V _ik

(36)

I=1 wherein, 2 ..., 30, ω is an inertia weight, is a constant between [0,1]; c ₁And c ₂Being learning rate, also is a non-negative constant; Rand ₁And rand ₂It is the random number that produces between [0,1]; V _Ik∈ [V _Max, V _Max], and V _MaxIt is the speed maximal value of an appointment.Can find out that from top two formula the moving direction of particle is by three part decisions, own original speed V _Ik, with the range difference (P of the own optimum position that experiences _Ik-X _Ik) and with the range difference (P of the optimum position of colony experience _Gk-X _Ik), and respectively by weight coefficient ω, c ₁And c ₂Determine its relative importance.The standard that iteration stops is the optimal-adaptive degree that perhaps reaches appointment according to maximum iteration time.

Step7: the individual extreme value of each particle is upgraded with following formula:

${P_i}^{\&prime;}=\left\{\begin{array}{cc}{X_i}^{\&prime;},\mathrm{if}& f\left({X_i}^{\&prime;}\right)\&GreaterEqual;f\left(P_i\right)\\ P_i,\mathrm{if}& f\left({X_i}^{\&prime;}\right)<f\left(P_i\right)\end{array}\right.$

(37)

Estimate each particle, calculate the fitness value of particle,, then upgrade this particle position if be better than the current individual extreme value of this particle.

Step8: the global extremum to all particles is chosen as follows:

P _g′＝P _imax′

(38)

P wherein _Imax' be the maximum Pi ' of adaptive value.

Step9: new particle more, whether check meets termination condition, if current iterations has reached predefined maximum times (or reach least error require), then stops iteration, the output optimum solution, otherwise continue more new particle.

The image information that the present invention can provide according to the user goes out many characteristic distances through many feature calculation of extracting image, utilizes particle cluster algorithm to optimize weighting parameter, and many characteristic distances are merged, and matees with the characteristics of image storehouse of appointment then.Thereby retrieve efficiently and the image of the image similarity that is retrieved, retrieval rate is high, has satisfied customer requirements.

Claims (2)

1. multi-feature image retrieval method based on particle cluster algorithm; It is characterized in that: the image information that this method provides according to the user; Many feature calculation through extracting image go out many characteristic distances, utilize particle cluster algorithm to optimize weighting parameter, and many characteristic distances are merged; Mate with the characteristics of image storehouse of appointment then, thereby retrieve efficiently and the image of the image similarity that is retrieved; Specifically comprise the steps:

Step 1: establish P _tBe the image that to retrieve, N width of cloth image arranged, P in the image data base _iBe i width of cloth image in the shape library, 1≤i≤N;

Step 2: extract a plurality of proper vectors of image that are retrieved, normalization, characteristic set is W={W ₁, W ₂..., W _K, wherein K is the characteristic number of extracting, K>=1;

Step 3: obtain P according to following formula _t, P _iAbout characteristic W _KCharacteristic distance D _i

$D(q,t)={\left(\underset{m=0}{\overset{M-1}{\mathrm{\&Sigma;}}}{|w_q\left(m\right)-w_t\left(m\right)|}^2\right)}^{\frac{1}{2}}$

(7)

Step 4: with distance B i normalization, obtain D according to following formula _i' mistake! Do not find Reference source.；

$d_i^{\&prime;}=\left\{\begin{array}{cc}\frac{d_i-\mathrm{mD}+3\mathrm{\&sigma;}}{6\mathrm{\&sigma;}}& d_i^{\&prime;}\&Element;(\mathrm{mD}-3\mathrm{\&sigma;},\mathrm{mD}+3\mathrm{\&sigma;})\\ 0& d_i^{\&prime;}\&le;\mathrm{mD}-3\mathrm{\&sigma;}\\ 1& d_i^{\&prime;}\&GreaterEqual;\mathrm{mD}-3\mathrm{\&sigma;}\end{array}\right.$

(8)

Wherein $\mathrm{MD}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{d}_i|n=\mathrm{1,2},...,N$

${\mathrm{\&sigma;}}^2=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{(d_i-\mathrm{mD})}^{2}n=\mathrm{1,2},...,N$

Step 5: K characteristic distance merged according to following formula;

D _Is=x ₁D ' ₁+ x ₂D ' ₂+ ... + x _KD ' _K, x wherein ₁..., x _K∈ [0,1], and x ₁+ ... + x _K=1; (9)

Utilize particle cluster algorithm to carry out parameter optimization, obtain optimized weights and distribute Xbest;

Step 6: after distributing Xbest according to optimized weights, obtain the comprehensive characteristics fusion distance, and retrieve, the result is fed back to the user, retrieve and the image of the image similarity that is retrieved according to image similarity.

2. the multi-feature image retrieval method based on particle cluster algorithm according to claim 1 is characterized in that: in the step 5, it is following to utilize particle cluster algorithm to carry out the step of parameter optimization:

Step1: suppose that search space is that K ties up and population has m particle, i particle represented the parameter vector X of a K dimension _i=(x _I1, x _I2..., x _IK), (i=1,2 ..., m), promptly i particle is X in the position of the search volume of K dimension _iIn other words, each particle position is the feasible solution that potential weights distribute, a random initializtion in allowed limits;

Step2: with X _iObjective function of substitution calculates its fitness, weighs good and bad according to the size of fitness; Objective function f (X _i) be made as many signature searchs result's under the current weight PVR index;

Step3: i the particle speed of " circling in the air " also is the vector of a K dimension, is expressed as: V _i=(v _I1, v _I2..., v _IK), (i=1,2 ..., m), random initializtion in allowed limits;

Step4: the optimal location note that i particle oneself searches is made P _i=(p _I1, p _I2..., p _IK), (i=1,2 ..., m); The P of each particle _iThe coordinate initialization is set to its current location, and calculates the fitness value of its corresponding individual extreme point, and promptly the weights of current individual representative distribute the PVR index result who retrieves;

Step5: the optimal location note that whole population searches is up to now made P _g=(p _G1, p _G2..., p _GK); P _gBe initialized as in the step best individual body position in all individual extreme values, be current best weight value distribution;

Step6: come particle is carried out iterative operation according to following formula:

V’ _ik＝ω·V _ik+c ₁·rand ₁·(P _ik-X _ik)+c ₂·rand ₂·(P _gk-X _ik)

(10)

X’ _ik＝X _ik+V _ik

(11)

I=1 wherein, 2 ..., m, ω is an inertia weight, is a constant between [0,1]; c ₁And c ₂Being learning rate, also is a non-negative constant; Rand ₁And rand ₂It is the random number that produces between [0,1]; V _Ik∈ [V _Max, V _Max], and V _MaxIt is the speed maximal value of an appointment;

Step7: the individual extreme value of each particle is upgraded with following formula:

${P_i}^{\&prime;}=\left\{\begin{array}{cc}{X_i}^{\&prime;},\mathrm{if}& f\left({X_i}^{\&prime;}\right)\&GreaterEqual;f\left(P_i\right)\\ P_i,\mathrm{if}& f\left({X_i}^{\&prime;}\right)<f\left(P_i\right)\end{array}\right.$

(12)

Estimate each particle, calculate the fitness value of particle,, then upgrade this particle position if be better than the current individual extreme value of this particle;

Step8: the global extremum to all particles is chosen as follows:

P _g′＝P _imax′

(13)

P wherein _Imax' be the maximum Pi ' of adaptive value;

Step9: new particle more, whether check meets termination condition, if current iterations has reached predefined maximum times or reached the least error requirement, then stops iteration, the output optimum solution, otherwise continue more new particle.

Images