CN1881211A - Graphic retrieve method - Google Patents

Abstract

The present invention relates to a method for image retrieval, and aims to provide a method for image retrieval that accurately judges image similarity and requires less storage data, a method for image retrieval, comprising: step 1, the user provides the retrieval request to the retrieval system; step 2 , the retrieval system automatically analyzes the requirement, obtains the feature corresponding to the requirement through feature extraction, and matches the feature with all corresponding features in the feature database; step 3, the retrieval system determines the retrieval result according to the feature matching result, and Provide the search result to the user. The invention has clear physical meaning, not only can distinguish the graphics of the same type according to the degree of similarity, but also can well distinguish different types of graphics; the required storage data is very little; the feature data has translation, rotation and scaling Invariance, at the same time, is more conducive to processing various types of images, and can be widely used in graphics and image processing.

Description

The method of graphic retrieve

Technical field

The present invention relates to the method for graphic retrieve, particularly use the method for carrying out graphic retrieve based on the boundary descriptor of direction distance distribution histogram.

Background technology

Along with science and technology development, the popularizing of computing machine and information network, numeral is obtained the popularization of equipment and digital storage equipment, makes the quantity of the various information datas that comprise literal, image, video, audio frequency etc. increase just with surprising rapidity.How to find from the data of magnanimity as soon as possible that Useful Information has become a very severe problem.Therefore, for the efficient access method of information data, particularly for the efficient access method of view data become more and more important.For this reason, people have proposed the CBIR method.This method can be by obtaining the content of image to the automatic analysis of characteristics of image, thereby opened up new approach for data access efficiently.At present, the CBIR method has been applied to medical treatment, textile printing and dyeing, meteorologic analysis, museum, library, release mechanism, film and television or the like industry-by-industry, there is QBIC (IBM1993) in wherein more famous system, Photobook (MIT1994), VisualSEEK (Columbia University1996) etc.CBIR has become a particular study field, although this field also is in preliminary developing stage, its achievement in research has been made contribution for the solution of many problems.More digital image retrieval system is also in development.

Characteristics algorithm is the core of graphic retrieve, and it is made up of feature extraction and two parts of characteristic matching.Feature extraction is meant by analyzing with the form of the data characteristic description with figure, and characteristic matching is meant the characteristic of different graphic is mated calculating to obtain the gap between the two.The different graphic features that characteristics algorithm extracted are different, and the quality of characteristics algorithm performance is directly determining the quality of the performance of search method.Therefore the core of search method finds a kind of efficiently characteristics algorithm fast exactly, makes the feature extraction of its existing efficient stable that fast accurate characteristic matching be arranged again.

Characteristics algorithm is mainly used in two places in retrieval.The one, the foundation of property data base.When setting up property data base, want the use characteristic algorithm that all figures are carried out feature extraction and obtain the characteristic of correspondence data, these characteristics and its graph of a correspondence are preserved by certain rule just set up property data base.The 2nd, by the automatic analysis of the requirement of user input.The user behind the tablet pattern, carries out feature extraction by characteristics algorithm to this figure in some way, and the feature that will obtain then in feature and the property data base is carried out characteristic matching, thereby finds the figure similar to tablet pattern.

The key issue of CBIR is to find a kind of not only stable but also search method efficiently.And shape is a kind of important images feature that is widely used in various search methods.Use shape analyze and the method for classifying a lot, have some new methods to propose in recent years again, constant moments method is wherein more typically arranged:

Square is a kind of statistical form to image, and its calculating will be used all relevant pixels in image or the zone.(x, y), if its segmentation is continuously and only non-vanishing on limited the point on XY plane, then provable its each rank square exists to a digital picture function f.F (x, p+q rank square y) is defined as:

$m_{\mathrm{pq}}=\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}{x}^p{y}^{q}f(x,y)$

F (x, p+q rank central moment y) is defined as:

${\mathrm{\μ}}_{\mathrm{pq}}=\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}{(x-\stackrel{\‾}{x})}^p{(y-\stackrel{\‾}{y})}^{q}f(x,y)$

X=m wherein ₁₀/ m ₀₀, y=m ₀₁/ m ₀₀, barycentric coordinates just.F (x, normalization central moment y) can be expressed as:

${\mathrm{\η}}_{\mathrm{pq}}=\frac{{\mathrm{\μ}}_{\mathrm{pq}}}{{\mathrm{\μ}}_{00}^{\mathrm{\γ}}}$

γ=(p+q)/2+1 wherein, p+q=2,3, L.

Some normalized second orders and third central moment made up just can obtain the constant square { φ of 7 translations, rotation and convergent-divergent image ₁, φ ₂, L, φ ₇.Every width of cloth image in the image library is all added up, and 7 invariant moments that obtain every width of cloth image are as characteristic, and 7 invariant moments with image to be checked compare then, just can judge the similarity of figure.

Though the characteristic of constant moments method is constant for translation, rotation and the convergent-divergent of image, but because this statistic does not have clear physical meaning, can accurately not be described, thereby its testing result is quite coarse, can not satisfies the requirement of data access efficiently figure.

Other relatively typically have the geometric parameter method, wavelet multiresolution rate descriptor index method, direction details histogram method, Edge Distance histogram method, local feature frequency method, tangent line space law, hidden Markov method based on the image retrieval of shape.The common deficiency of these methods is: its description to the shape similarity is felt and can not be conformed to well with human similarity, can not obtain the retrieval effectiveness that people wish in many cases.For example, said method even can not sort out to basic configurations such as typical triangle and quadrilaterals.

Summary of the invention

The objective of the invention is at above-mentioned the deficiencies in the prior art, a kind of figure similarity accuracy of judgement is provided and requires the few figure retrieving method of storage data.

For achieving the above object, the present invention proposes a kind of method of graphic retrieve, specifically carry out following steps:

Step 1, user will retrieve and require to offer searching system;

Step 2, described searching system are analyzed described requirement automatically, obtain this by feature extraction and require characteristic of correspondence, and all individual features in this feature and the property data base are mated;

Step 3, described searching system be according to characteristic matching deterministic retrieval result as a result, and provide the user with result for retrieval.

Wherein said step 2 is specially:

The direction distance distribution histogram of step 21, calculating graph outline;

Step 22, from described direction distance distribution histogram, detect the spike number of described profile;

The descriptor of step 23, the described profile of structure also carries out matching operation.

The concrete following steps of carrying out of wherein said step 21:

Step 210, detect the edge of described figure;

The center of gravity of step 211, calculating figure;

Step 212, the profile of described figure is scanned;

The center of gravity of the profile of step 213, the described figure of calculating;

Step 214, the profile to described figure scans again, obtains a distance distribution histogram;

Step 215, described distance distribution histogram is carried out normalization;

Step 216, from described normalized distance distribution histogram, find out minimum value and value and maximal value, to peaked direction described distance distribution histogram is sorted, obtain the direction distance distribution histogram according to minimum value.

Wherein the described scanning of step 212 and step 214 for from 0 spend to 359 the degree scannings of totally 360 directions, sweep spacing be 1 the degree.

The concrete following steps of carrying out of described step 22:

Step 220, described direction distance distribution histogram is carried out gaussian filtering, eliminate the noise on the described contour curve, obtain a new direction distance distribution histogram;

Step 221, described new direction distance distribution histogram is carried out first derivation, obtain the first order derivative histogram;

Step 222, described first order derivative histogram is carried out differentiate, obtain the second derivative histogram;

Step 223, find out in all second derivatives, calculate and preserve the number of these values greater than the value of a threshold value.

Wherein the number of the value described in the step 223 is exactly the spike number of profile.

First feature of described profile spike number as outline used in matching operation described in the step 23, uses the supplemental characteristic of the direction distance distribution histogram of described profile as outline.

The present invention has the following advantages:

(1) physical significance distinctness, not only the figure of same kind inside can be made a distinction according to similarity degree, also can well different types of graphical demarcation be come, be better than the prior art scheme in this, can not be because the prior art scheme only possesses the ability distinguished by similarity degree with pattern classification, this means that also using this programme to carry out graphic retrieve can obtain more accurate result for retrieval;

(2) storage data of the presently claimed invention seldom, every width of cloth graph image only needs 13 characteristics to be described, be in same magnitude with the prior art scheme, concerning a database with n width of cloth graph image, the complexity of the calculating of each retrieval is O (mn), wherein m＜＜n, retrieval rate is fast, is fit to portable terminal and uses;

(3) characteristic of the present invention has translation, rotation and convergent-divergent unchangeability, has kept the advantage of prior art scheme, simultaneously, more helps handling various dissimilar images, and range of application is wider.

Below by drawings and Examples, technical scheme of the present invention is described in further detail.

Description of drawings

Fig. 1 is the synoptic diagram of the embodiment intermediate cam shape image of figure retrieving method of the present invention;

Fig. 2 is the process flow diagram of the embodiment of figure retrieving method of the present invention;

Fig. 3 is edge-detected image synoptic diagram among the embodiment of figure retrieving method of the present invention;

Fig. 4 is the distance distribution histogram among the embodiment of figure retrieving method of the present invention;

Fig. 5 is the normalized distance distribution histogram among the embodiment of figure retrieving method of the present invention;

Fig. 6 is the direction distance distribution histogram among the embodiment of figure retrieving method of the present invention;

The direction distance distribution histogram of Fig. 7 after for the gaussian filtering among the embodiment of figure retrieving method of the present invention;

The distance distribution histogram of Fig. 8 after for the first derivation among the embodiment of figure retrieving method of the present invention;

The distance distribution histogram of Fig. 9 after for the second order differentiate among the embodiment of figure retrieving method of the present invention;

Figure 10 detects synoptic diagram for the impulse among the embodiment of figure retrieving method of the present invention.

Embodiment

Triangular image with M*N is an example below, specifically describes technical scheme of the present invention, and this triangular image as shown in Figure 1.

As described in Figure 2, the process flow diagram for the embodiment of figure retrieving method of the present invention specifically may further comprise the steps:

Step 100, user will be retrieved by inquiry modes such as illustration, sketch, camera inputs and require to offer searching system (the illustration mode: the user specifies a width of cloth to retrieve from existing sample graphics; The sketch mode: the instrument that user's using system the provides sketch that draws is retrieved; The camera mode: the user uses camera to take needs the object of retrieval to retrieve); It is a triangle that the retrieval here requires;

Step 101, use edge detection method to detect the edge of triangular pattern, obtain edge image f (x, y), f ∈ (0,1) here, x ∈ (1, L, N), y ∈ (1, L, M); Image after the rim detection is as shown in Figure 3:

Step 102, calculate the center of gravity O=(O of triangular pattern according to formula (1) _x, O _y):

$O_x=\frac{\underset{y=1}{\overset{N}{\mathrm{\Σ}}}\underset{x=1}{\overset{M}{\mathrm{\Σ}}}\mathrm{xf}(x,y)}{\underset{y=1}{\overset{N}{\mathrm{\Σ}}}\underset{x=1}{\overset{M}{\mathrm{\Σ}}}f(x,y)},O_y=\frac{\underset{y=1}{\overset{N}{\mathrm{\Σ}}}\underset{x=1}{\overset{M}{\mathrm{\Σ}}}\mathrm{yf}(x,y)}{\underset{y=1}{\overset{N}{\mathrm{\Σ}}}\underset{x=1}{\overset{M}{\mathrm{\Σ}}}f(x,y)}---\left(1\right)$

Step 103, begin the profile of figure is begun to carry out the scanning of totally 360 directions of 0-359 degree from the edge image center of gravity, sweep spacing is 1 degree, finds on each direction from center of gravity edge farthest, be the position of profile, obtain contour images g (x, y), here g ∈ (0,1), and x ∈ (1, L, N), y ∈ (1, L, M);

Step 104, calculate the center of gravity of profile, obtain centre of gravity place C=(C according to formula (2) _x, C _y):

$C_x=\frac{\underset{y=1}{\overset{N}{\mathrm{\Σ}}}\underset{x=1}{\overset{M}{\mathrm{\Σ}}}\mathrm{xg}(x,y)}{\underset{y=1}{\overset{N}{\mathrm{\Σ}}}\underset{x=1}{\overset{M}{\mathrm{\Σ}}}g(x,y)},C_y=\frac{\underset{y=1}{\overset{N}{\mathrm{\Σ}}}\underset{x=1}{\overset{M}{\mathrm{\Σ}}}\mathrm{yg}(x,y)}{\underset{y=1}{\overset{N}{\mathrm{\Σ}}}\underset{x=1}{\overset{M}{\mathrm{\Σ}}}g(x,y)}---\left(2\right);$

The profile centre of gravity place that step 105, basis obtain carries out the scanning of 0-359 degree again, and calculates the distance from the center of gravity to the profile on each direction, thereby obtains the distance distribution histogram of one 360 dimension $H_D=(h_0^D,h_1^D,L,h_{359}^D),$ This distance distribution histogram verifies that easily this histogram remains unchanged for the translation of image as shown in Figure 4;

Step 106, according to formula (3) to H _DCarry out normalization, obtain normalized distance distribution histogram

$h_N=(h_0^N,h_1^N,L,h_{359}^N);$

$h_i^N=\frac{h_i^D-\mathrm{Min}}{\mathrm{Max}-\mathrm{Min}},i=\mathrm{0,1},L,359.---\left(3\right)$

Max wherein, Min is respectively histogram H _DIn maximal value and minimum value; Normalization can make histogram that the convergent-divergent of image is remained unchanged, and can guarantee normalized distance distribution histogram H thus _NTranslation, convergent-divergent to image are constant; As shown in Figure 5, be the normalized distance distribution histogram of triangle;

Step 107, from normalized distance distribution histogram H _NIn find all minimum value (be assumed to P) and maximal value (it is individual to be assumed to Q), it is right so just can to obtain P * Q minimum value-maximal value.Here we to define minimum value-peaked distance poor for both call numbers, that is:

$\mathrm{Dist}(h_i^{\max },h_{i^{\&prime;}}^{\min })=\left\{\begin{array}{c}|i-i^{\&prime;}|,\mathrm{if}|i-i^{\&prime;}|<180\\ 359-|i-i^{\&prime;}|,\mathrm{else}.\end{array}\right.$

Find the shortest that of distance a pair of from these minimum value-maximal value centering (if having a plurality of, can continue comparison maximal value-minimum value other of indicated zone are worth carry out the shortest right the searching of minimum value-maximal value of distance), and be first element with minimum value, guarantee that simultaneously maximal value is in preceding 180 elements, to H _NAccording to minimum value-maximal value direction rearrangement, obtain a new distance distribution histogram, i.e. the direction distance distribution histogram $H_R=(h_0^R,h_1^R,L,h_{359}^R),$ As shown in Figure 6; Why can do like this, be because scanning is spent from O to 359, histogrammic first that obtains is adjacent on graph outline with last, so just can make in the histogram in any two zones that remain on 180 degree at interval, thereby can obtain direction distance distribution histogram recited above by circulation; By direction distance distribution histogram H _RConstruction process as can be seen its rotation to image remain unchanged; Therefore, direction distance distribution histogram H _RTranslation, rotation, convergent-divergent to image all remain unchanged; This histogram is quantized according to formula (4), obtain the characteristic B=(b of 12 elements ₀, b ₁, L, b ₁₁).

$b_i=\frac{1}{30}\underset{j=i\&times;30}{\overset{(i+1)\&times;30-1}{\mathrm{\&Sigma;}}}{h}_j^R,i=\mathrm{0,1},L,11.---\left(4\right);$

Obtained normalized direction distance distribution histogram H above _RThis histogram is the one dimension mapping of X-Y scheme profile, spike on the spike of profile and the direction distance distribution histogram is exactly one to one so, so can detect by the impulse to the histogrammic second derivative of profile, the spike number in the profile is detected;

Step 108, to H _RCarry out gaussian filtering, the noise on the contour elimination curve obtains a more level and smooth direction distance distribution histogram $H_G=(h_0^G,h_0^G,L,h_{359}^G),$ As shown in Figure 7;

Step 109, to histogram H _GCarry out first derivation, obtain the first order derivative histogram $H_F=(h_0^F,h_1^F,L,h_{359}^F),$ As shown in Figure 8;

Step 110, again to histogram H _FCarry out differentiate, obtain the second derivative histogram $H_B=(h_0^B,h_1^B,L,h_{359}^B),$ As shown in Figure 9;

Step 111, find in all second derivatives value greater than a certain threshold value, i.e. impulse, the number of calculating and preserving these impulses, just the number of spike is as eigenwert A; Impulse detects as shown in figure 10;

Step 112, by characteristic A that obtains previously and B, can structural configuration descriptor, i.e. boundary descriptor: D={A|B};

Matching operation for boundary descriptor, at first to consider from the similarity of figure, because the similarity of figure does not also have perfect mathematical definition now, so can only on the basis of experiment, go to approach people's similarity sensation to a certain extent, here, because the profile similarity of figure can be described the similarity of figure preferably, and the spike on the profile is the principal character of outline, so use the principal character of profile spike number as outline here, the direction distance distribution histogram of profile is as the outline supplemental characteristic; Main matching characteristic is used for shape is distinguished into different classifications, and the auxiliary matched feature is used for the shape of the inside of same type is distinguished; Because the difference between the classification is greater than the difference between the type inside certainly, so for any two boundary descriptor D ₁={ A ₁| B ₁, D ₂={ A ₂| B ₂}: if A ₁≠ A ₂, then require A ₁With A ₂Between minor increment for D ₁With D ₂Between the influence of distance be greater than and equal B ₁With B ₂Between ultimate range for D ₁With D ₂Between the influence of distance, promptly the difference of principal character is always greater than the difference of supplemental characteristic, then D ₁With D ₂Matching formula is:

$\mathrm{Dist}\_D=\frac{360}{361}\mathrm{Dist}\_A+\frac{1}{361}\mathrm{Dist}\_B---\left(5\right)$

Wherein:

$\mathrm{Dist}\_A=\frac{1}{360}|A_1-A_2|,$

$\mathrm{Dist}\_B=\sqrt{\frac{1}{12}\underset{i=0}{\overset{11}{\mathrm{\&Sigma;}}}{(b_{1,i}-b_{2,i})}^2,}$

Dist_D，Dist_A，Dist_B∈[0，1]；

Step 113, system offer the user with the mode of the pairing figure of characteristic matching result by Pagination Display.

The present invention is textural the direction distance distribution histogram, also can sort according to maximum-minimum direction or other directions, thereby obtain the same distance distribution histogram with directivity; In addition, also can be in the detection of spike, the detection method of other types such as use small echo obtains the number of spike; Can also be at boundary descriptor textural, the data that the direction distance distribution histogram are quantified as other numbers such as 6,8,10 are constructed similar boundary descriptor, also can use L ₁, L ₂Other distance calculation formula such as norm calculate the similarity between figure.

It should be noted last that, above embodiment is only unrestricted in order to technical scheme of the present invention to be described, although the present invention is had been described in detail with reference to preferred embodiment, those of ordinary skill in the art is to be understood that, can make amendment or be equal to replacement technical scheme of the present invention, and not break away from the spirit and scope of technical solution of the present invention.

Claims (6)

1, a kind of method of graphic retrieve, wherein carry out following steps:

Step 1, user will retrieve and require to offer searching system;

Step 2, described searching system are analyzed described requirement automatically, obtain this by feature extraction and require characteristic of correspondence, and all individual features in this feature and the property data base are mated;

Step 3, described searching system be according to characteristic matching deterministic retrieval result as a result, and provide the user with result for retrieval.

2, the method for graphic retrieve according to claim 1, wherein said step 2 is specially:

The direction distance distribution histogram of step 21, calculating graph outline;

Step 22, from described direction distance distribution histogram, detect the spike number of described profile;

The descriptor of step 23, the described profile of structure also carries out matching operation.

3, the method for graphic retrieve according to claim 2, the concrete following steps of carrying out of wherein said step 21:

Step 210, detect the edge of described figure;

The center of gravity of step 211, calculating figure;

Step 212, the profile of described figure is scanned;

The center of gravity of the profile of step 213, the described figure of calculating;

Step 214, the profile to described figure scans again, obtains a distance distribution histogram;

Step 215, described distance distribution histogram is carried out normalization;

Step 216, from described normalized distance distribution histogram, find out minimum value and value and maximal value, to peaked direction described distance distribution histogram is sorted, obtain the direction distance distribution histogram according to minimum value.

4, the method for graphic retrieve according to claim 3, wherein described in step 212 and the step 214 scanning be specially: from 0 spend to 359 the degree totally 360 directions scan, sweep spacing be 1 the degree.

5, according to the method for claim 2 or 3 or 4 described graphic retrieves, wherein said step 22 is specifically carried out following steps:

Step 220, described direction distance distribution histogram is carried out gaussian filtering, eliminate the noise on the described contour curve, obtain a new direction distance distribution histogram;

Step 221, described new direction distance distribution histogram is carried out first derivation, obtain the first order derivative histogram;

Step 222, described first order derivative histogram is carried out differentiate, obtain the second derivative histogram;

Step 223, find out in all second derivatives greater than the value of a threshold value, calculate and preserve the number of these values, the number of described value is exactly the spike number of profile.

6, according to the method for claim 2,3 or 4 described graphic retrieves, wherein first feature of described profile spike number as outline used in matching operation described in the step 23, uses the supplemental characteristic of the direction distance distribution histogram of described profile as outline.

Images