CN106202250A - Class-based comprehensive characteristics and the image search method of complete class coupling - Google Patents

Abstract

The invention discloses a kind of class-based comprehensive characteristics and the image search method of complete class coupling, it is to easily cause the loss of information for respective regions matching process in region based CBIR, complete area matching process is computationally intensive, the problem of similar area repeated matching, the class-based image search method proposed, being the complete representative image of energy and several classes with uniqueness region clusterings all in image, the level in class extracts feature and mates；The class in image is obtained by accelerating statistical regions merging and neighbour's transmission method；The comprehensive characteristics method utilizing class carries out feature extraction to class；Complete class matching process is used to be its distribution weights according to each class importance in the picture and mate.Testing Corel 1000 and Caltech 256 image library, result shows that the inventive method has more preferable retrieval effectiveness than traditional search method based on region.

Description

Class-based comprehensive characteristics and the image search method of complete class coupling

Technical field

The invention belongs to image retrieval technologies field, be specifically related to a kind of class-based comprehensive characteristics and complete class coupling Image search method.

Background technology

Along with the development of computer networking technology, information is more and more carried by digital picture, how in quantity Huge image library retrieves people's image interested and becomes a popular research field, CBIR (Content-based Image Retrieval, CBIR) technology receives much concern due to its good retrieval effectiveness.Although Technology is in many aspects compared with text based image retrieval (Text-based Image Retrieval, TBIR) for CBIR technology Tool has great advantage, but is also faced with some problems, the most importantly difference between lower-level vision feature and high-level semantics, at present Solve this problem most common method be region based CBIR (Region-based Image Retrieval, RBIR) technology.RBIR is by being divided into some regions image and carrying out Region Matching and retrieve image, and this more meets regarding of people Feel sensory perceptual system.Traditional RBIR system segments the image into region by partitioning algorithm, then mates to come real to region Existing image retrieval.How to select the key issue that the region participating in coupling is always in RBIR research.

In research in early days, occur in that some use the system of respective regions coupling, such as Berkeley university The NeTra of Blobword and California university.These systems need manually to region, feature, and feature power equivalence selects Select, this ensure that the region participating in coupling is consistent with the interest of user, and different Search Requirements can be met, but user is frequent Can be for selecting which region and which type of parameter can reach optimum search effect and perplex.

The achievement in research that another moderns attractes attention is general area coupling (the Integrated region that Li et al. proposes Matching, IRM), the method allows all regions in image to participate in coupling, it is ensured that the integrity of image information, reduces Split the inaccurate impact on retrieval result.IRM has initially obtained application also in the SIMPLIcity system of Stanford University Achieving good effect, up to now, IRM is still widely used as a classical matching process.

In some searching systems based on respective regions coupling, the region participating in coupling can be selected by system automatically Select.Such as irregular interest region (Irregular regions of interest, the IROI) search method proposed at Yuber In, system selects the some regions closest with point of interest to participate in mating according to the position feature in each region, it is not necessary to artificial Select, but such system of selection is easily omitted some key areas and caused the loss of image information.

There is document (Yang, X.H., Cai, L.J.: ' Adaptive region matching for region-based image retrieval by constructing region importance index’,IET Comput.Vis., 2013,8, (2), pp.1 11) propose a kind of complete area matching process (Active region matching, ARM), can Switch between respective regions coupling and complete area coupling with the feature according to region.Semantic region (Semantic in document Meaningful region, SMR) whether method met threshold value judged this district by area and the position feature of zoning Whether territory participates in coupling, when all regions are satisfied by condition SMR condition, and now system is complete area coupling, is otherwise indivedual Region Matching.

In RBIR, artificial selection's matching area can increase the burden of user；Based on regional location area features automatic Matching area selects easily to omit key area and cause information dropout；Complete area match party rule has region repeated matching Drawback, especially when the region of repeated matching is background can to retrieval result have an immense impact on.Therefore, in RBIR how Select participate in coupling region be research emphasis be also difficult point.

Summary of the invention

It is an object of the invention to overcome above-mentioned problems of the prior art, propose a kind of class-based image retrieval side Method, is ensureing that image information avoids repeated matching while complete, is can complete representative image region clusterings all in image And there are several classes of uniqueness, the level in class extracts feature and mates.

The technical scheme is that class-based comprehensive characteristics and the image search method of complete class coupling, the method It is a kind of new image retrieval framework of the aspect of class from image, comprises the steps:

Step 1: extract the class in image

Propose to accelerate statistical regions to merge and neighbour propagation (Accelerated statistical region merging And affinity propagation, ASRM-AP) method extracts the class in image: first passes through ASRM method and image entered Row segmentation, carries out color and vein feature extraction to segmentation gained region, then carries out AP cluster and finds out and have similar spy in image The region levied is marked the class obtaining in image；

Step 2: extract the comprehensive characteristics of class

Comprehensive characteristics (Integrated feature of category, the IFOC) method utilizing class carries out table to class State, in IFOC, the color of class, textural characteristics are combined as the comprehensive spy of class with the quantity in apoplexy due to endogenous wind region, distribution characteristics Levy；The color of class and textural characteristics are respectively by color histogram and local binary patterns (the Local Binary in apoplexy due to endogenous wind region Pattern, LBP) method obtains, and the quantative attribute of class is obtained by the quantity in statistics apoplexy due to endogenous wind region being normalized, distribution It is characterized as that apoplexy due to endogenous wind region is gone up in the picture, in, the distribution histogram that lower floor is three layers；

Step 3: class is carried out complete match

Propose towards class on the basis of complete area coupling (Integrated region matching, IRM) algorithm Complete class coupling (Integrated category matching, ICM) method, and replace area to be class right of distribution with barycenter Value, obtains the distance between image, it is achieved image retrieval.

Acceleration statistical regions described in above-mentioned steps 1 merges and neighbour propagates (Accelerated statistical Region merging and affinity propagation, ASRM-AP) method specifically includes following steps:

Step 2.1: image I is carried out ASRM segmentation.First image is carried out packing process, i.e. image is divided into 3 × 3 Block, the pixel average of computing block is as I_SA pixel value obtain figure I_S, then to I_SCarry out statistical regions merging (Statistical region merging, SRM) is split, then segmentation result is mapped to original image I；

ASRM partitioning algorithm:

(1) image I is divided into the block b of 3 × 3, using each piece of packing as scheming I_SIn a pixel, pixel Value v is the pixel average of block b；

$v_a=\underset{i=1}{\overset{9}{\mathrm{\Σ}}}{b}_a\left(i\right)/9,a\∈(R,G,B)---\left(1\right)$

Wherein, three Color Channels during R, G, B are image；

(2) to figure I_SCarry out SRM segmentation；

(3) I_SSegmentation result be mapped to image I, obtain the segmentation result of I；

Step 2.2: ASRM segmentation gained region is carried out AP cluster, obtains the class in image.AP clustering algorithm is by disappearing Breath circulation way progressively determines cluster centre, and namely iteration updates Attraction Degree matrix R=[r (i, k)] and degree of membership matrix A =[a (i, k)], finally realize high-quality self-adaption cluster；Its more new regulation is as follows:

1) Attraction Degree matrix R is updated with degree of membership matrix A and similarity matrix S=[s (i, k)]:

$r(i,k)=s(i,k)-\underset{k^{\′}\≠k}{max}\{a(i,k^{\′})+s(i,k^{\′})\}---\left(2\right)$

2) by Attraction Degree matrix R renewal degree of membership matrix A:

$a(i,k)=\min \{0,r(k,k)+\underset{i^{\′}\∉\{i,k\}}{\mathrm{\Σ}}\max (0,r(i^{\′},k\left)\right)\}---\left(3\right)$

$a(k,k)=\underset{i^{\′}\≠k}{\mathrm{\Σ}}\max \{0,r(i^{\′},k)\}---\left(4\right)$

Wherein, any two region during i and k is ASRM segmentation gained region；(i k) represents the k Attraction Degree to i to r；a(i, K) i degree of membership to k is represented；Certain region that i ' is non-i, certain region that k ' is non-k；S (i, k) is the similarity of i to k:

S (i, k)=-| | c_ri-c_rk||²-||t_ri-t_rk||²

(5)

$c_r(h,s,v)=\frac{N(h,s,v)}{N_{total}}---\left(6\right)$

$t_r\left(lbp\right)=\frac{N\left(lbp\right)}{N_{total}}---\left(7\right)$

$lbp=\underset{r=0}{\overset{7}{\Σ}}l(g_r-g_c)2^r---\left(8\right)$

$l\left(x\right)=\left\{\begin{array}{cc}1,& x\&GreaterEqual;0\\ 0,& x<0\end{array}\right.---\left(9\right)$

c_ri,c_rkAnd t_ri,t_rkRepresent the color and vein characteristic vector of i and k respectively, respectively by statistical color histogram drawn game Portion's binary pattern (Local Binary Pattern, LBP) method is tried to achieve, and sees formula (6), formula (7)；N (h, s, v) Representative Region H in territory, the value of S, V Color Channel corresponds to h, the number of pixel when s, v；N_totalFor number of pixels in region；N (lbp) is district In territory, LBP value is the number of pixels of lbp；g_c, g_rIt is respectively pixel average and the pixel value of position r of 3 × 3 pieces；

When i with k is equal, being configured s by deflection parameter p, the biggest data k of p are the most likely chosen as in cluster The heart:

S (k, k)=p*mean (s (k :))

(10)

Function mean (s (k :)) be s (k: the average of element in)；S (k :) represent region k and other all regions Similarity；Parameter p takes 0.6.

After AP cluster, according to cluster result, region is marked, all pixels in kth apoplexy due to endogenous wind region is set to k, Finally give the labelling figure that pixel is 1 to n, thus obtain the class of image.

The comprehensive characteristics IFOC method utilizing class proposed in above-mentioned steps 2 calculates the comprehensive characteristics of class；It is below The region quantity of class and Regional Distribution Characteristics extracting method in IFOC, comprise the following steps:

Region quantity feature n of (a) class_i:

n_i=g (N)

(11)

Wherein, N represents that the number in the i-th class region after ASRM-AP, g (x) expression are normalized computing to variable x, Its rule is:

$g\left(x\right)=\left\{\begin{array}{cc}(x-1)/4,& x\&le;5\\ 1,& otherwise\end{array}\right.---\left(12\right)$

Homogeneous region number is normalized to [0,1], when number is more than 5, it is believed that number of regions is more, and character pair value is 1。

The Regional Distribution Characteristics of (b) class:

Owing to the object in image often exists certain regularity of distribution at vertical direction, as sky is often positioned on image Layer, animal, trees etc. are often positioned in image middle level, therefore divide the image into upper strata, middle level, lower floor, and the district to each apoplexy due to endogenous wind Level belonging to the barycenter of territory is added up, and obtains the Regional Distribution Characteristics l of class_i:

l_i=[N_h(i)/N,N_m(i)/N,N_l(i)/N]

(13)

Wherein, N represents the number in the i-th apoplexy due to endogenous wind region, N in the image obtained by ASRM-AP algorithm_h(i), N_m(i), N_l(i) Represent the i-th class region number in image upper strata, middle level and lower floor respectively；

The comprehensive characteristics of class is f:

f_i=[c_i,t_i,n_i,l_i]

(14)

$c_i=\underset{k=1}{\overset{n}{\&Sigma;}}{a}_k{c}_{rk}/n---\left(15\right)$

$t_i=\underset{k=1}{\overset{n}{\&Sigma;}}{a}_k{t}_{rk}/n---\left(16\right)$

Wherein c_i, t_i, n_i, l_iRepresent the color of i-th class, texture, the region quantity of class and Regional Distribution Characteristics, n respectively For the number in i-th apoplexy due to endogenous wind region, a_kThe ratio of the number of pixels of place class is accounted for for the number of pixels of region k.c_rk,t_rkFor region The color and vein feature of k, in formula (6), formula is given in (7)；

The ICM class matching process proposed in above-mentioned steps 3, replaces area to be that class distributes weights with barycenter, obtains between image Distance；Image I₁With image I₂In class use C respectively₁=(ca₁,ca₂,…,ca_m), C₂=(ca '₁,ca′₂,…,ca′_n) table Show, then the distance of two images is D (I₁, I₂):

$D(I_1,I_2)=\underset{i,j}{\mathrm{\&Sigma;}}{d}_{i,j}{s}_{i,j}---\left(17\right)$

d_i,j=α₁|c_i-c′_j|+α₂|t_i-t′_j|+α₃|n_i-n′_j|+α₄|l_i-l′_j|

(18)

Wherein d_i,jRepresent class ca_iWith ca '_jDistance；s_i,jRepresent region ca_iWith ca '_jThe interest-degree of coupling, interest-degree square Battle array S is:

$S=\left[\begin{array}{cccc}{s}_{1,1}& s_{1,2}& ...& s_{1,n}\\ s_{2,1}& s_{2,2}& ...& s_{2,n}\\ \&CenterDot;& \&CenterDot;& & \&CenterDot;\\ \&CenterDot;& \&CenterDot;& & \&CenterDot;\\ \&CenterDot;& \&CenterDot;& & \&CenterDot;\\ s_{m,1}& s_{m,2}& ...& s_{m,n}\end{array}\right]---\left(19\right)$

c_i,ti,n_i,l_iWith c '_j,t′_j,n′_j,l′_jRepresent image I respectively₁In i-th class and image I₂In jth class Color, texture, apoplexy due to endogenous wind region quantity and the characteristic vector of area distribution, computational methods are respectively in formula (15), (16), (11) (13) be given in；α₁,α₂,α₃,α₄For the weights of different characteristic, α₁+α₂+α₃+α₄=1 and α₁,α₂,α₃,α₄∈ (0,1), this 0.25 it is disposed as in bright；

Image I is carried out canny rim detection, formula (22) obtains the barycenter (x of gained texture maps_I,y_I), according to often The region of individual apoplexy due to endogenous wind is to mean value computation its interest-degree s of centroid distance_i,j:

s_i,j=max (s_i,s_j)

(20)

$s_i=1/\left(\underset{r\&Element;{\mathrm{ca}}_i}{\mathrm{\&Sigma;}}\right|\left|f\right(r)-(x_I,y_I\left)\right||/N_i)---\left(21\right)$

$f\left(r\right)=\&lsqb;\underset{x_k\&Element;r(x,y)}{\&Sigma;}{x}_k/N,\underset{y_k\&Element;r(x,y)}{\&Sigma;}{y}_k/N\&rsqb;---\left(22\right)$

Wherein, f_rFor region r being sought the function of barycenter；N_iIt it is class ca_iThe number in middle region；x_kAnd y_kIt is respectively in the r of region The abscissa of pixel and vertical coordinate；N is the number of pixel in the r of region；s_jWith s_iComputational methods are similar to.

Beneficial effects of the present invention: the present invention proposes a kind of class-based image search method, is ensureing image information Avoid repeated matching while Wan Zheng, be can complete representative image and there are the some of uniqueness region clusterings all in image Individual class, the level in class extracts feature and mates.The invention have the advantages that

1, propose class-based image retrieval (Category-based Image Retrieval, CaBIR) framework, find out All of class in image, carries out feature extraction and class coupling, obtains the distance of image class.The information in RBIR system that solves is lost The problem of repeated matching of becoming estranged.

2, propose to accelerate statistical regions to merge and neighbour propagation (Accelerated statistical region Merging and affinity propagation, ASRM-AP) method is in order to obtain the class in image.SRM be a kind of based on The dividing method of region growing, it is ensured that the integrity in region, is accelerated SRM processing to improve system effectiveness simultaneously； Utilize adaptive AP clustering algorithm similar area to be assembled and is labeled as different classes, adjacent similar area is carried out simultaneously Merge, to prevent over-segmentation.

3, proposing comprehensive characteristics (Integrated feature of category, the IFOC) method of class, IFOC is class The visual informations such as the color in middle region, texture and the quantity in apoplexy due to endogenous wind region and distributed intelligence are combined as the feature of class.IFOC side Method enriches the information that feature comprises, and reduces the difference of low-level features and high-level semantics.

4, propose complete class coupling (Integrated category matching, ICM) method each class is distributed not Same weights also carry out mating to obtain image distance.Compared with the complete area coupling relying on area ratio to be region distribution weights (Integrated region matching, IRM), ICM method is that class distributes weights by the barycenter in apoplexy due to endogenous wind region, more meets The visually-perceptible of people.

Below with reference to accompanying drawing, the present invention is described in further details.

Accompanying drawing explanation

Fig. 1 is the algorithm flow chart of the present invention；

Fig. 2 is the extraction algorithm flow chart of ASRM-AP class；

Fig. 3 is the Regional Distribution Characteristics vector that in image I, three class regions are corresponding；

Fig. 4 is segmentation and the classification results of parts of images；

The variation instance figure of precision ratio when Fig. 5 is to use IFOC and Conventional visual feature；

Fig. 6 is snow mountain and the segmentation of basketry image and classification results；

Fig. 7 is the three kinds of methods be given average precisions in Corel-1000 and Caltech-256 image library；

Fig. 8 (a) is when Corel-1000, the average precision broken line graph of each method during different K values；

Fig. 8 (b) is when Caltech-256, the average precision broken line graph of each method during different K values；

Fig. 9 is the retrieval result of parts of images.

Detailed description of the invention

In the present invention, it is proposed that CaBIR retrieves framework, extract the class in image by ASRM-AP method, utilize IFOC pair Class carries out the feature extraction difference with reduction low-level features with high-level semantics, and ICM method is used to according to each class in the picture Importance be its distribution weights mating.By region clusterings all in image for the complete representative image of energy and are had Several classes of uniqueness, the level in class extracts feature and mates, and avoids repetition while guarantee information is complete Coupling, thus improve retrieval quality.Idiographic flow is presented in Fig. 1.

Present invention step is as follows:

1. extract the class in image

Class is the set of similar area in image, and in the present invention, the method taking segmentation and cluster to combine extracts image In class.First pass through ASRM method image is split, segmentation gained region is carried out color and vein feature extraction, then Carry out AP cluster to find out the region in image with similar features and as same class and be marked.Owing to AP algorithm is self adaptation Clustering algorithm, system can determine the number of class in image according to picture material.

Retrieval may be towards large-scale image storehouse, and therefore dividing method should be simple efficient.SRM method is a kind of base In the dividing method of region growing, its target is region image being divided into and having following two feature, i.e. in arbitrary region The value of all pixels of each passage and the difference of the average of this passage；The pixel of an at least passage in arbitrary region The difference of average of value and this passage of neighborhood not in certain threshold value.

SRM algorithm idiographic flow is:

(1) all unduplicated four connected pixels are found out in image I to [(x₁,y₁),(x₂,y₂)], wherein (x₁,y₁), (x₂,y₂) it is the coordinate of two pixels of pixel centering.

(2) calculate value f of pixel pair, and according to the value ascending order of f be pixel to arrange region index matrix S (often goes in S Coordinate for pixel pair).

F=max (R (x₁,y₁)-R(x₂,y₂),G(x₁,y₁)-G(x₂,y₂),B(x₁,y₁)-B(x₂,y₂))

(1)

Wherein, three Color Channels during R, G, B are image I.

(3) judge whether the region (prime area is pixel itself) belonging to two pixels in often going in S meets Anticipation function P.

$P(R,R^{\&prime;})=\left\{\begin{array}{cc}yes,& {\max }_{a\&Element;(R,G,B)}|\stackrel{\&OverBar;}{R_a^{\&prime;}}-\stackrel{\&OverBar;}{R_a}|\&le;\sqrt{b^2\left(R\right)+b^2\left(R^{\&prime;}\right)}\\ no,& otherwise\end{array}\right.---\left(2\right)$

$b\left(R\right)=g\sqrt{\frac{1}{2Q\left|R\right|}\ln \frac{\left|R\right|}{\mathrm{\&delta;}}}---\left(3\right)$

Wherein, R and R ' represents the affiliated area of two pixels in S row,WithRepresent two region R to be judged The average of Color Channel a middle with R '.G is the value of Color Channel resolution, and parameter Q determines the complexity of segmentation, and value is divided the most greatly The number cut is the most.| R | is the pixel number in the R of region.δ is the maximum of probability of P (R, R ')=no, and default value is the least.This G=256 in invention, Q=20, δ=1/ (6 | I |²)。

(4) traveling through index matrix S from top to bottom, it may be judged whether meet formula 2, if met, merging the two region, If be unsatisfactory for, next line is judged.

Due to SRM algorithm need to be averaged in two corresponding regions by pixel each in image, anticipation function is sentenced Breaking and wait computing, the time of operation is longer.SRM is improved, before segmentation, image is carried out packing process, i.e. image is split It is the block of 3 × 3, the value that the pixel average of computing block is wrapped as this, then packing image is carried out SRM segmentation, then segmentation Result is mapped to original image.

It is accelerated SRM algorithm processing:

(1) image I is divided into the block b of 3 × 3, using each piece of packing as scheming I_SIn a pixel, pixel Value v is the pixel average of block b.

$v_a=\underset{i=1}{\overset{9}{\mathrm{\&Sigma;}}}{b}_a\left(i\right)/9,a\&Element;(R,G,B)---\left(4\right)$

(2) to figure I_SCarry out SRM segmentation.

(3) I_SSegmentation result be mapped to image I, obtain the segmentation result of I.

SRM operand is represented by:

$cal=L\left(I\right)H\left(I\right)\sqrt{\mathrm{\&tau;}L\left(I\right)H\left(I\right)}---\left(5\right)$

Wherein L (I), H (I) are respectively length and the height of image；τ is constant and no more than 1；It is accelerated SRM processing, I_S Length and high be respectively L₁(I)=L (I)/3, H₁(I)=H (I)/3, the operand cal after acceleration process₁For:

$\begin{array}{c}{\mathrm{cal}}_1=\left(L\right(I)/3)\left(H\right(I)/3)\sqrt{\mathrm{\&tau;}\left(L\right(I)/3)\left(H\right(I)/3)}\\ =L\left(I\right)H\left(I\right)\sqrt{\mathrm{\&tau;}L\left(I\right)H\left(I\right)}/27\\ =cal/27\end{array}---\left(26\right)$

The operation time after acceleration in theory processes that understands is about original 1/27.

AP clustering algorithm progressively determines cluster centre by message circulation way, and namely iteration updates Attraction Degree matrix R=[r (i, k)] and degree of membership matrix A=[a (i, k)], finally realize high-quality self-adaption cluster.More new regulation is as follows:

(1) Attraction Degree matrix R is updated with degree of membership matrix A and similarity matrix S=[s (i, k)]:

$r(i,k)=s(i,k)-\underset{k^{\&prime;}\&NotEqual;k}{max}\{a(i,k^{\&prime;})+s(i,k^{\&prime;})\}---\left(7\right)$

(2) by Attraction Degree matrix R renewal degree of membership matrix A:

$a(i,k)=\min \{0,r(k,k)+\underset{i^{\&prime;}\&NotElement;\{i,k\}}{\mathrm{\&Sigma;}}\max (0,r(i^{\&prime;},k\left)\right)\}---\left(8\right)$

$a(k,k)=\underset{i^{\&prime;}\&NotEqual;k}{\mathrm{\&Sigma;}}\max \{0,r(i^{\&prime;},k)\}---\left(9\right)$

Wherein, i and k is any two region in ASRM segmentation gained region, and (i k) represents the k Attraction Degree to i to r；a(i, K) i degree of membership to k is represented；Certain object that i ' is non-i, certain object that k ' is non-k；S (i, k) is the similarity of i to k:

S (i, k)=-| | c_ri-c_rk||²-||t_ri-t_rk||²

(10)

$c_r(h,s,v)=\frac{N(h,s,v)}{N_{total}}---\left(11\right)$

$t_r\left(lbp\right)=\frac{N\left(lbp\right)}{N_{total}}---\left(12\right)$

$lbp=\underset{r=0}{\overset{7}{\&Sigma;}}l(g_r-g_c)2^r---\left(13\right)$

$l\left(x\right)=\left\{\begin{array}{cc}1,& x\&GreaterEqual;0\\ 0,& x<0\end{array}\right.---\left(14\right)$

c_ri,c_rkAnd t_ri,t_rkRepresent the color and vein characteristic vector of i and k respectively, respectively by statistical color histogram drawn game Portion's binary pattern (Local Binary Pattern, LBP) method is tried to achieve, and sees formula 11, formula 12；(h, s v) represent region to N Middle H, the value of S, V Color Channel corresponds to h, the number of pixel when s, v；N_totalFor number of pixels in region；N (lbp) is region Middle LBP value is the number of pixels of lbp；g_c, g_rIt is respectively pixel average and the value of position r of 3 × 3 pieces；

When i with k is equal, being configured s by deflection parameter p, the biggest data k of p are the most likely chosen as in cluster The heart:

S (k, k)=p*mean (s (k :))

(15)

Function mean (s (k :)) be s (k: the average of element in)；S (k :) represent region k and other all regions Similarity；Parameter p takes 0.6.

After AP cluster, according to cluster result, region is marked, all pixels in kth apoplexy due to endogenous wind region is set to k, Finally give the labelling figure that pixel is 1 to n, thus obtain the class in image.Idiographic flow is as shown in Figure 2.Fig. 2 is class in image Extraction flow chart 1. in image I 3 × 3 pieces I that pack to obtain_S, 2. to I_SCarry out SRM segmentation, 3. I_SSegmentation result is mapped to I, 4. SRM segmentation result is carried out AP cluster the class obtaining in image to similar area labelling, 5. same apoplexy due to endogenous wind adjacent area is entered Row merges.

2. extract the comprehensive characteristics of class

Image search method in the present invention is class-based method, compared to utilizing field color, stricture of vagina in RBIR system Region is stated by the visual signatures such as reason, shape, proposes a kind of IFOC method and states class.In IFOC, class The quantity in color, textural characteristics and apoplexy due to endogenous wind region, distribution characteristics combine the comprehensive characteristics as class, the method for this multiple features Reduce the difference between low-level features and high-level semantics.The color of class and textural characteristics are by the color and vein feature in apoplexy due to endogenous wind region (obtaining when clustering region AP) average represents, mainly introduces region quantity and the Regional Distribution Characteristics extracting method of class.

(1) region quantity feature n of class_i:

n_i=g (N)

(16)

Wherein, N represents that the number in the i-th class region after ASRM-AP, g (x) expression are normalized computing to x, its rule It is then:

$g\left(x\right)=\left\{\begin{array}{cc}(x-1)/4,& x\&le;5\\ 1,& otherwise\end{array}\right.---\left(17\right)$

Homogeneous region number is normalized to [0,1], when number is more than 5, it is believed that number of regions is more, and character pair value is 1。

(2) Regional Distribution Characteristics of class:

Owing to the object in image often exists certain regularity of distribution at vertical direction, as sky is often positioned on image Layer, animal, trees etc. are often positioned in image middle level.Therefore upper strata, middle level, lower floor, and the district to each apoplexy due to endogenous wind are divided the image into Level belonging to the barycenter of territory is added up, and obtains the Regional Distribution Characteristics l of class_i:

l_i=[N_h(i)/N,N_m(i)/N,N_l(i)/N]

(18)

Wherein, N represents the number in the i-th apoplexy due to endogenous wind region, N in the image obtained by ASRM-AP algorithm_h(i), N_m(i), N_l(i) Represent the i-th class region number in image upper strata, middle level and lower floor respectively.

The comprehensive characteristics of class is f:

f_i=[c_i,t_i,n_i,l_i]

(19)

$c_i=\underset{k=1}{\overset{n}{\&Sigma;}}{a}_k{c}_{rk}/n---\left(20\right)$

$t_i=\underset{k=1}{\overset{n}{\&Sigma;}}{a}_k{t}_{rk}/n---\left(21\right)$

Wherein c_i, t_i, n_i, l_iRepresent the color of i-th class, texture, the region quantity of class and Regional Distribution Characteristics, n respectively For the number in i-th apoplexy due to endogenous wind region, a_kThe ratio of the number of pixels of place class is accounted for for the number of pixels of region k.c_rk,t_rkFor region The color and vein feature of k, in formula (11), formula is given in (12)；

3. pair class carries out complete match

IRM algorithm allows a region to mate with multiple regions, reduces the error that segmentation inaccuracy causes, improves The robustness of system.But, in IRM, the weights in region are only determined by region area, bigger when containing area in image Background area such as sky, meadow etc., can produce larger interference to retrieval.The present invention proposes the ICM towards class on the basis of IRM Matching process, replaces area to be the class distribution weights participating in coupling with barycenter, more meets the visually-perceptible of people.Image I₁And image I₂In class C₁=(ca₁,ca₂,…,ca_m), C₂=(ca '₁,ca′₂,…,ca′_n) represent, the distance of two images is D (I₁, I₂):

$D(I_1,I_2)=\underset{i,j}{\mathrm{\&Sigma;}}{d}_{i,j}{s}_{i,j}---\left(22\right)$

d_i,j=α₁|c_i-c′_j|+α₂|t_i-t′_j|+α₃|n_i-n′_j|+α₄|l_i-l′_j|

(23)

Wherein d_i,jRepresent class ca_iWith ca '_jDistance；s_i,jRepresent region ca_iWith ca '_jThe interest-degree of coupling, interest-degree square Battle array S is:

$S=\left[\begin{array}{cccc}{s}_{1,1}& s_{1,2}& ...& s_{1,n}\\ s_{2,1}& s_{2,2}& ...& s_{2,n}\\ \&CenterDot;& \&CenterDot;& & \&CenterDot;\\ \&CenterDot;& \&CenterDot;& & \&CenterDot;\\ \&CenterDot;& \&CenterDot;& & \&CenterDot;\\ s_{m,1}& s_{m,2}& ...& s_{m,n}\end{array}\right]---\left(24\right)$

c_i,t_i,n_i,l_iWith c '_j,t′_j,n′_j,l′_jRepresent image I respectively₁In i-th class and image I₂In jth class Color, texture, apoplexy due to endogenous wind region quantity and the characteristic vector of area distribution, computational methods are respectively in formula 20,21,16 and 18 Be given；α₁,α₂,α₃,α₄For the weights of different characteristic, α₁+α₂+α₃+α₄=1 and α₁,α₂,α₃,α₄∈ (0,1), is all provided with in the present invention It is set to 0.25.

Article (XIA Dingyuan, FU Pian, LIU Liduan. " Improved image retrieval Algorithm for integrated region matching ", [J] .CEA, 2012,48 (26): 197-200.) middle proposition A kind of region based on center interest-degree computational methods, and obtained good effect.But, the nearlyest not Representative Region of distance center Territory is more paid close attention to by people, and in image, texture is complicated, and the region that gradient is big is often easier to attract people to note, it is proposed that a kind of Interest-degree computational methods based on centroid distance.Image I is carried out canny rim detection, formula 27 obtains gained texture maps Barycenter (x_I,y_I), according to mean value computation its interest-degree s in the region of each apoplexy due to endogenous wind to centroid distance_i,j:

s_i,j=max (s_i,s_j)

(25)

$s_i=1/\left(\underset{r\&Element;{\mathrm{ca}}_i}{\mathrm{\&Sigma;}}\right|\left|f\right(r)-(x_I,y_I\left)\right||/N_i)---\left(26\right)$

$f\left(r\right)=\&lsqb;\underset{x_k\&Element;r(x,y)}{\&Sigma;}{x}_k/N,\underset{y_k\&Element;r(x,y)}{\&Sigma;}{y}_k/N\&rsqb;---\left(27\right)$

Wherein, f_rFor region r being sought the function of barycenter；n_iIt it is class ca_iThe number in middle region；x_kAnd y_kIt is respectively in the r of region The abscissa of pixel and vertical coordinate；N is the number of pixel in the r of region；s_jWith s_iComputational methods are similar to.

Experimental result and analysis

1, experimental situation and the image library of employing

The system test environment of experiment is: Duo i5CPU, 3.20GHz, 8.0GB RAM；Windows 7 operating system； Matlab R2014a develops software.Use retrieval image library Core-1000 the most frequently used in CBIR experiment Image library and Caltech-256 image library.The former comprises 10 class images, respectively primitive man, sandy beach, traces, bus, probably Dragon, elephant, horse, flower, snow mountain, food, every class comprise 100 big little be the jpeg image of 256 × 384 and 384 × 256.The latter Then containing 30607 images of 256 class difference objects, every apoplexy due to endogenous wind comprises 80 to 827 images, experiment have selected rifle, National flag of United States, both shoulders bag, baseball glove, hoop, bat, bathtub, stein, Vespertilio, motorboat 10 class totally 1299 images are carried out Retrieval.

2, performance estimating method

Evaluation criterion the most frequently used in CBIR system is precision ratio and recall ratio, and precision ratio represents inspection The image number relevant to query example figure obtained of rope and the ratio retrieving gained total number of images；Recall ratio represents to be retrieved The ratio that the image number relevant to query example figure arrived is total with associated picture.Precision ratio and the highest explanation of value of recall ratio The effect of algorithm is the best, i.e. systematic function is the best.But precision ratio and recall ratio are often again contradictory relation.Standard is looked in order to improve Rate and reduce the image set of return, inevitably result in recall ratio decline；Otherwise, in order to improve recall ratio, increase the image returned Number, is also easily included more uncorrelated image result, thus causes precision ratio to decline.Therefore, the present invention only selects With precision ratio P, retrieval result is evaluated.

P=n_k/K,

(25)

K is retrieval fruiting quantities, n_kFor the number of associated picture in retrieval result.The average precision of algorithm is:

$\stackrel{\&OverBar;}{P}=\underset{q=1}{\overset{i}{\&Sigma;}}{P}_q/i,---\left(26\right)$

P_qRepresent q width query example figure precision ratio, i is the number of query example figure.

3, about image segmentation and the experiment of classification

In order to verify the feasibility of ASRM, from Corel-100 and Caltech-256 image library, respectively randomly select 100 width These images are carried out using the CaBIR retrieval of SRM and using the CaBIR retrieval of ASRM, and it are average to calculate system by image respectively Operation time and precision ratio.Result is as shown in table 1, in Corel-1000 image library, SRM is accelerated system and averagely runs Time shortens 26 times, and precision ratio declines 0.03 percentage point；The system that in Caltech-256 image library is accelerated SRM is put down Time of all running shortens 22 times, and precision ratio declines 0.04 percentage point.Therefore, ASRM dividing method is used can to look into standard with minimum Rate is the operation time that cost is greatly shortened system.

Table 1 uses the CaBIR system operation time of SRM and ASRM to contrast with precision ratio

Fig. 4 gives the classification results of parts of images.Image a is divided into background, dinosaur, soil three class, image b quilt Being divided into palm fibre horse, Baima, meadow three class, image c is divided into vegetable, cookies, white tablecloth, red cake four class, image c quilt Being divided into background, the metal body of a gun, the wooden body of a gun three class, image e is divided into star, background, white stripes, red streak four Class.As can be seen from the figure ASRM-AP algorithm adaptive can determine the number comprising class in image, and accurately similar Region is divided into same class.

4, about the experiment of IFOC feature

In IFOC, system adds the distribution characteristics to apoplexy due to endogenous wind region and the identification ability of quantative attribute.Selected part Image carries out contrast experiment to traditional characteristic and IFOC feature.In Fig. 5, when a1, b1, c1, d1 are to use Conventional visual feature Part retrieval result, part retrieval result when a2, b2, c2, d2 are to use IFOC feature.A1, although retrieving result in a2 It is horse, but a1 comprises the image of two dry goods, image that a2 comprises a dry goods and ranking behind；5th the 6th width in b1 Owing to containing some classes similar to query graph vision and by false retrieval in image；C, d two groups also there is analogue.Logical Cross contrast it is found that introduce Regional Distribution Characteristics can effectively strengthen the system separating capacity to class, only full when similar Foot vision phase Sihe distribution could obtain relatively low matching distance time similar, thus improves precision ratio.Fig. 5 uses IFOC and biography The change of precision ratio during system visual signature.A, b, c, d are query example figure；When a1, b1, c1, d1 are for using Conventional visual feature Front 6 width retrieval results；Front 6 width retrieval results when a2, b2, c2, d2 are to use IFOC feature.

5, about the experiment of interest-degree computational methods based on barycenter

In Fig. 6, obtain the class in snow mountain and basketry image by ASRM-AP algorithm.Table 2 gives employing distinct methods and obtains The interest-degree of the every class arrived and the contrast of subjective interest-degree.Subjective interest-degree is that inhomogeneity is beaten by 10 people according to interest level The average divided.From Table 2, it can be seen that in snow mountain and basketry image method gained interest-degree based on barycenter and subjective interest Spend closest to, next to that method based on center, interest-degree based on area differs maximum with subjective interest-degree.Thus may be used Knowing, interest-degree computational methods based on barycenter more meet human vision property.

The contrast of interest-degree method asked by 2 three kinds of table

Fig. 7 gives three kinds of methods average precision in Corel-1000 and Caltech-256 image library, tests table Obtain the highest precision ratio during bright employing interest-degree based on barycenter computational methods, can preferably reflect that people is to inhomogeneous emerging Interest degree.

6 CaBIR and the contrast of additive method

The class-based image search method that the present invention is proposed and SIMPLIcity method SRM-IRM method, MN-MIN These five kinds of region based CBIR methods of method, SIS method and MN-ARM contrast.Table 3 table 4 sets forth K= Each system precision ratio in Corel-1000 and Caltech-256 image library when 20, Fig. 8 gives each system during different K values The variation tendency of average precision.

In table 3, the precision ratio of CaBIR falls behind MN-ARM method at African, building, mountain three apoplexy due to endogenous wind.This be due to original inhabitant, Building, these objects of mountain want complexity to be unfavorable for the segmentation of image compared with objects such as horse, flower, sandy beach, dinosaur, buses, adopt for this With complete class matching process, to reduce the impact that segmentation error is brought.In other class images, CaBIR precision ratio is higher than its other party Method, especially in horse and bus image, precision ratio improves about 5%.Generally speaking, the inventive method average precision is 77.19%, higher than the 76.60% of MN-ARM method.

In table 4, the inventive method precision ratio in most of images is all obviously improved, especially at national flag of United States image In, precision ratio exceeds 25% than second；In basketry image, precision ratio falls behind MN-ARM method 0.37%, and this is due to mostly In number basketry image, net is little and sparse, is easily left in the basket, thus causes the loss of class during segmentation.On the whole at Caltech- In 256 image libraries, the inventive method average precision exceeds 5.68% than MN-ARM, significantly improves.The inventive method exists Advantage in Corel-1000 image library is not the highest in Caltech-256 image library, is due to the object in the latter Do not have the former complicated, it is easier to carry out splitting and classifying, the most more can embody the advantage of class-based image retrieval.

During table 3 K=20, in Corel-1000 image library, the inventive method contrasts with additive method precision ratio

During table 4 K=20, in Caltech-256 image library, the inventive method contrasts with additive method precision ratio

Fig. 8 is the average precision broken line graph of distinct methods, sets forth distinct methods K value be 20,40,60,80, Average precision when 100.As can be seen from the figure along with the increase of K value, precision ratio the most constantly declines, but the inventive method Being consistently higher than other search methods based on region, again demonstrate the advantage of CaBIR, Fig. 9 gives the retrieval of parts of images As a result, the first two ten width in flower, horse, bus, national flag of United States, both shoulders bag, six kinds of query example figure retrieval results of bat are related to Image.

Conclusion: in the present invention, it is proposed that a kind of class-based image search method.First, obtained by ASRM-AP method Class in image, then utilizes IFOC class to carry out feature extraction to reduce the difference of low-level features and high-level semantics, ICM method Being used to according to each class importance in the picture is that its distribution weights carrying out mate and obtain the similarity of image.By handle In image, all region clusterings are the complete representative image of energy and several classes with uniqueness, and the level in class extracts feature also Mate, while guarantee information is complete, avoids repeated matching, thus improves retrieval quality.Test result indicate that this Bright method has more preferable effect than existing region based CBIR method.Further work center of gravity is to improve further How the accuracy of segmentation, make this method have the more preferable suitability in the image of complex background, and improve as far as possible The efficiency of system.

The present invention is ensureing that image information avoids repeated matching while complete in sum, and regions all in image are gathered Collection is the complete representative image of energy and several classes with uniqueness, and the level in class extracts feature and mates.The present invention Advantage include:

1, propose class-based image retrieval (Category-based Image Retrieval, CaBIR) framework, find out All of class in image, carries out feature extraction and class coupling, obtains the distance of image class.The information in RBIR system that solves is lost The problem of repeated matching of becoming estranged.

2, propose to accelerate statistical regions to merge and neighbour propagation (Accelerated statistical region Merging and affinity propagation, ASRM-AP) method is in order to obtain the class in image.SRM be a kind of based on The dividing method of region growing, it is ensured that the integrity in region, is accelerated SRM processing to improve system effectiveness simultaneously； Utilize adaptive AP clustering algorithm similar area to be assembled and is labeled as different classes, adjacent similar area is carried out simultaneously Merge, to prevent over-segmentation.

3, proposing comprehensive characteristics (Integrated feature of category, the IFOC) method of class, IFOC is class The visual informations such as the color in middle region, texture and the quantity in apoplexy due to endogenous wind region and distributed intelligence are combined as the feature of class.IFOC side Method enriches the information that feature comprises, and reduces the difference of low-level features and high-level semantics.

4, propose complete class coupling (Integrated category matching, ICM) method each class is distributed not Same weights also carry out mating to obtain image distance.Compared with the complete area coupling relying on area ratio to be region distribution weights (Integrated region matching, IRM), ICM method is that class distributes weights by the barycenter in apoplexy due to endogenous wind region, more meets The visually-perceptible of people.

In present embodiment, the part of not narration in detail belongs to the known conventional means of the industry, chats the most one by one State.Exemplified as above is only the illustration to the present invention, is not intended that the restriction to protection scope of the present invention, every and basis Invent within same or analogous design belongs to protection scope of the present invention.

Claims (4)

1. The most class-based comprehensive characteristics and the image search method of complete class coupling, it is characterised in that the method is a kind of from figure The image retrieval framework that in Xiang, the aspect of class is set out, comprises the steps:

   Step 1: extract the class in image

   Propagate ASRM-AP method extract the class in image by accelerating statistical regions merging and neighbour: first pass through ASRM method Image is split, segmentation gained region is carried out color and vein feature extraction, then carry out AP cluster and find out tool in image The region having similar features is marked the class obtaining in image；

   Step 2: extract the comprehensive characteristics of class

   Class is stated by the comprehensive characteristics IFOC method utilizing class, in IFOC, the color of class, textural characteristics and apoplexy due to endogenous wind district The quantity in territory, distribution characteristics combine the comprehensive characteristics as class；The color of class and textural characteristics are respectively by the face in apoplexy due to endogenous wind region Color Histogram and local binary patterns LBP method obtain, and the quantative attribute of class is by adding up the quantity in apoplexy due to endogenous wind region and returning One change obtains, and distribution characteristics is that apoplexy due to endogenous wind region is gone up in the picture, in, the distribution histogram that lower floor is three layers；

   Step 3: class is carried out complete match

   On the basis of complete area coupling IRM algorithm, propose the complete class towards class mate ICM method, and replace area with barycenter Distribute weights for class, obtain the distance between image, it is achieved image retrieval.
2. Class-based comprehensive characteristics the most as claimed in claim 1 and the image search method of complete class coupling, it is characterised in that Acceleration statistical regions described in step 1 merges and neighbour propagates ASRM-AP method and specifically includes following steps:

   Step 2.1: image I is carried out ASRM segmentation: first image is carried out packing process, i.e. image is divided into 3 × 3 Block, the pixel average of computing block is as I_SA pixel value obtain figure I_S, then to I_SCarry out statistical regions and merge SRM segmentation, Again segmentation result is mapped to original image I；

   ASRM partitioning algorithm:

   (1) image I is divided into the block b of 3 × 3, using each piece of packing as scheming I_SIn a pixel, value v of pixel is The pixel average of block b；

   $v_a=\underset{i=1}{\overset{9}{\mathrm{\&Sigma;}}}{b}_a\left(i\right)/9,a\&Element;(R,G,B)---\left(1\right)$

   Wherein, three Color Channels during R, G, B are image；

   (2) to figure I_SCarry out SRM segmentation；

   (3) I_SSegmentation result be mapped to image I, obtain the segmentation result of I；

   Step 2.2: ASRM segmentation gained region is carried out AP cluster, obtains the class in image；AP clustering algorithm is passed by message Broadcast mode progressively determines cluster centre, and namely iteration updates Attraction Degree matrix R=[r (i, k)] and degree of membership matrix A=[a (i, k)], finally realize high-quality self-adaption cluster；Its more new regulation is as follows:

   1) Attraction Degree matrix R is updated with degree of membership matrix A and similarity matrix S=[s (i, k)]:

   $r(i,k)=s(i,k)-\underset{k^{\&prime;}\&NotEqual;k}{max}\{a(i,k^{\&prime;})+s(i,k^{\&prime;})\}---\left(2\right)$

   2) by Attraction Degree matrix R renewal degree of membership matrix A:

   $a(i,k)=\min \{0,r(k,k)+\underset{i^{\&prime;}\&NotElement;\{i,k\}}{\mathrm{\&Sigma;}}\max (0,r(i^{\&prime;},k\left)\right)\}---\left(3\right)$

   $a(k,k)=\underset{i^{\&prime;}\&NotEqual;k}{\mathrm{\&Sigma;}}\max \{0,r(i^{\&prime;},k)\}---\left(4\right)$

   Wherein, any two region during i and k is ASRM segmentation gained region；(i k) represents the k Attraction Degree to i to r；A (i, k) table Show the i degree of membership to k；Certain region that i ' is non-i, certain region that k ' is non-k；S (i, k) is the similarity of i to k:

   S (i, k)=-| | c_ri-c_rk||²-||t_ri-t_rk||² (5)

   $c_r(h,s,v)=\frac{N(h,s,v)}{N_{total}}---\left(6\right)$

   $t_r\left(lbp\right)=\frac{N\left(lbp\right)}{N_{total}}---\left(7\right)$

   $lbp=\underset{r=0}{\overset{7}{\&Sigma;}}l(g_r-g_c)2^r---\left(8\right)$

   $l\left(x\right)=\left\{\begin{array}{cc}1,& x\&GreaterEqual;0\\ 0,& x<0\end{array}\right.---\left(9\right)$

   c_ri, c_rkAnd t_ri, t_rkRepresent the color and vein characteristic vector of i and k respectively, respectively by statistical color histogram and local two Binarization mode LBP method is tried to achieve, and sees formula (6), formula (7)；(h, s, v) represent H in region to N, and the value of S, V Color Channel corresponds to The number of pixel when h, s, v；N_totalFor number of pixels in region；N (lbp) is that in region, LBP value is the number of pixels of lbp；g_c, g_rIt is respectively pixel average and the pixel value of position r of 3 × 3 pieces；

   When i with k is equal, being configured s by deflection parameter p, the biggest data k of p are the most likely chosen as cluster centre:

   S (k, k)=p*mean (s (k :)) (10)

   Function mean (s (k :)) be s (k: the average of element in)；It is similar to other all regions that s (k :) represents region k Degree；Parameter p takes 0.6；

   After AP cluster, according to cluster result, region is marked, all pixels in kth apoplexy due to endogenous wind region are set to k, finally Obtain the labelling figure that pixel is 1 to n, thus obtain the class of image.
3. Class-based comprehensive characteristics the most as claimed in claim 1 and the image search method of complete class coupling, it is characterised in that Step 2 proposes a kind of comprehensive characteristics IFOC method utilizing class and calculates the comprehensive characteristics of class；It it is below the district of class in IFOC Territory quantity and Regional Distribution Characteristics extracting method, comprise the following steps:

   Region quantity feature n of (a) class_i:

   n_i=g (N) (11)

   Wherein, N represents that the number in the i-th class region after ASRM-AP, g (x) expression are normalized computing to variable x, its rule It is then:

   $g\left(x\right)=\left\{\begin{array}{cc}(x-1)/4,& x\&le;5\\ 1,& otherwise\end{array}\right.---\left(12\right)$

   Homogeneous region number is normalized to [0,1], when number is more than 5, it is believed that number of regions is more, and character pair value is 1；

   The Regional Distribution Characteristics of (b) class:

   Owing to the object in image often exists certain regularity of distribution at vertical direction, as sky is often positioned in image upper strata, Animal, trees etc. are often positioned in image middle level, therefore divide the image into upper strata, middle level, lower floor, and the regional quality to each apoplexy due to endogenous wind Belonging to the heart, level is added up, and obtains the Regional Distribution Characteristics l of class_i:

   l_i=[N_h(i)/N, N_m(i)/N, N_l(i)/N] (13)

   Wherein, N represents the number in the i-th apoplexy due to endogenous wind region, N in the image obtained by ASRM-AP algorithm_h(i), N_m(i), N_l(i) difference Represent the i-th class region number in image upper strata, middle level and lower floor；

   The comprehensive characteristics of class is f:

   f_i=[c_i, t_i, n_i, l_i] (14)

   $c_i=\underset{k=1}{\overset{n}{\&Sigma;}}{a}_k{c}_{rk}/n---\left(15\right)$

   $t_i=\underset{k=1}{\overset{n}{\&Sigma;}}{a}_k{t}_{rk}/n---\left(16\right)$

   Wherein c_i, t_i, n_i, l_iRepresenting the color of i-th class respectively, texture, the region quantity of class and Regional Distribution Characteristics, n is i-th The number in individual apoplexy due to endogenous wind region, a_kThe ratio of the number of pixels of place class is accounted for for the number of pixels of region k；c_rk, t_rkFace for region k Color textural characteristics, in formula (6), formula is given in (7).
4. Class-based comprehensive characteristics the most as claimed in claim 1 and the image search method of complete class coupling, it is characterised in that The ICM class matching process proposed in step 3, replaces area to be that class distributes weights with barycenter, obtains the distance between image；Image I₁ With image I₂In class use C respectively₁=(ca₁, ca₂..., ca_m), C₂=(ca '₁, ca '₂..., ca '_n) represent, then two images Distance is D (I₁, I₂):

   $D(I_1,I_2)=\underset{i,j}{\mathrm{\&Sigma;}}{d}_{i,j}{s}_{i,j}---\left(17\right)$

   d_{I, j}=α₁|c_i-c′_j|+α₂|t_i-t′_j|+α₃|n_i-n′_j|+α₄|l_i-l′_j| (18)

   Wherein d_{I, j}Represent class ca_iWith ca '_jDistance；s_{I, j}Represent region ca_iWith ca '_jThe interest-degree of coupling, interest-degree matrix S For:

   $S=\left[\begin{array}{cccc}{s}_{1,1}& s_{1,2}& ...& s_{1,n}\\ s_{2,1}& s_{2,2}& ...& s_{2,n}\\ \&CenterDot;& \&CenterDot;& & \&CenterDot;\\ \&CenterDot;& \&CenterDot;& & \&CenterDot;\\ \&CenterDot;& \&CenterDot;& & \&CenterDot;\\ s_{m,1}& s_{m,2}& ...& s_{m,n}\end{array}\right]---\left(19\right)$

   c_i,t_i,n_i,l_iWith c '_j,t′_j,n′_j,l′_jRepresent image I respectively₁In i-th class and image I₂In the face of jth class Color, texture, apoplexy due to endogenous wind region quantity and the characteristic vector of area distribution, computational methods respectively in formula (15), (16), (11) and (13) be given in；α₁,α₂,α₃,α₄For the weights of different characteristic, α₁+α₂+α₃+α₄=1 and α₁,α₂,α₃,α₄∈ (0,1), the present invention In be disposed as 0.25；

   Image I is carried out canny rim detection, formula (22) obtains the barycenter (x of gained texture maps_I,y_I), according to each class In region to mean value computation its interest-degree s of centroid distance_i,j:

   s_i,j=max (s_i,s_j) (20)

   $s_i=1/\left(\underset{r\&Element;{\mathrm{ca}}_i}{\mathrm{\&Sigma;}}\right|\left|f\right(r)-(x_I,y_I\left)\right||/N_i)---\left(21\right)$

   $f\left(r\right)=\&lsqb;\underset{x_k\&Element;r(x,y)}{\&Sigma;}{x}_k/N,\underset{y_k\&Element;r(x,y)}{\&Sigma;}{y}_k/N\&rsqb;---\left(22\right)$

   Wherein, f_rFor region r being sought the function of barycenter；N_iIt it is class ca_iThe number in middle region；x_kAnd y_kIt is respectively pixel in the r of region Abscissa and vertical coordinate；N is the number of pixel in the r of region；s_jWith s_iComputational methods are similar to.