CN103473288B - A kind of image search method describing son based on mixing micro structure - Google Patents

Abstract

The present invention relates to a kind of image search method describing son based on mixing micro structure, first with mixing micro structure, son (Hybrid MSD is described, HMSD) the characteristics of image storehouse that image library is corresponding is generated, the content characteristic of retrieval image is extracted followed by HMSD, then similarity measurement criterion is utilized to measure, finally by similarity measurement sort result, and correspond to image and present to user.Son is described containing the direction under two kinds of conventional color space model and the micro structure of color owing to the present invention mixing micro structure, therefore in color, texture, shape facility and distribution of color information, have the strongest resolution capability, reflect the attribute of human visual system to a certain extent.The image retrieval performance of the image search method therefore describing son based on mixing micro structure has bigger lifting.

Description

A kind of image search method describing son based on mixing micro structure

Technical field

The present invention relates to a kind of image search method describing son based on mixing micro structure.

Background technology

CBIR (Content Based Image Retrieval, CBIR) basic thought is according to inspection Visually-perceptible content and semantic content that rope image is comprised set up characteristic vector, and then according to the similarity of characteristic vector Carry out coupling retrieval.Vision content feature conventional in CBIR mainly includes color of image, shape and texture etc..Face Color, shape, texture picture engraving content from different perspectives.In order to improve retrieval performance, need more reasonably description figure As content, the most comprehensive dissimilar feature is an up retrieving the effective ways of performance.Document " Image retrieval Based on micro-structure descriptor, Pattern Recognition, 2011,44 (9): 2123-2133. " disclose A kind of micro structure for feature extraction describes son (Micro-Structure Descriptor, MSD).The method utilizes micro- Structure has carried out effective comprehensive description to color, texture, shape and the distributed intelligence thereof of image.First will input certainly So image is transformed into hsv color space, then carries out the extraction of edge direction at the image obtained, then in edge side On image, extract micro structure figure, finally, reflect according on micro structure figure hsv color spatial image after quantization Penetrate, obtain final microstructure image, and microstructure image is carried out co-occurrence matrix and histogrammic description.Although should The characteristics of image such as method color combining, texture and shape, but owing to microstructure image is by edge direction image and face Color quantized image carries out seeking common ground what computing produced, and therefore edge directional information and colouring information descriptive power are not enough, and And microstructure analysis process is complicated in the method.

In sum, existing feature extracting method can not the color of effectively comprehensive characterization image, texture, edge The image content informations such as direction.

Summary of the invention

Solve the technical problem that

In place of the deficiencies in the prior art, the present invention proposes a kind of image describing son based on mixing micro structure Search method, overcomes the shortcoming of feature description scarce capacity in conventional images search method, improves the precision of image retrieval.

Technical scheme

A kind of image search method describing son based on mixing micro structure, it is characterised in that step is as follows:

Step 1: to images all in image library, utilizes mixing micro structure to describe the sub feature carrying out picture material and retouches State, generate corresponding characteristics of image storehouse；

Step 2: utilize mixing micro structure to describe son and extract the content characteristic of retrieval image；

Step 3: the characteristic vector in retrieval characteristics of image and characteristics of image storehouse is carried out similarity measurement；

Step 4: returned by image corresponding for the top n result of similarity measurement, similarity quantization method uses city City's block distance；

It is as follows that the mixing micro structure of described step 1 and step 2 describes sub-step:

Step a: input picture is respectively converted into the data under RGB and hsv color spatial model；

Step b: the view data under each color spatial model is carried out respectively edge direction extraction and color quantizing, Obtain color-quantized images and edge direction image；

Step c: respectively each color-quantized images and edge direction image carried out microstructure analysis and carry out Nogata Figure describes；

Step d: use co-occurrence matrix and rectangular histogram to enter the micro structure of two same types under different colours model Line description；

Step e: by color-quantized images micro structure rectangular histogram and the edge direction micro structure rectangular histogram of mixing of mixing Carry out comprehensive.

Edge direction extraction step in described step b is as follows:

Step (a): use Sobel operator to obtain each passage of image in x and y direction in three-dimensional color space Direction gradient；

Step (b): according to formulaObtain Formula a b=X_hX_v+Y_hY_v+Z_hZ_v, then according to angle formulae between vectorCalculate image Edge direction image；

Wherein, a (X_h,Y_h,Z_h) and b (X_v,Y_v,Z_v) represent gradient both horizontally and vertically, X respectively_hRepresent X Passage gradient in the horizontal direction, X_vRepresent the X passage gradient in vertical direction.Y_h、Y_vAnd Z_h、Z_vBe respectively Y, Z passage is in the gradient of both direction；

Step (c): the edge direction θ unified quantization of each pixel is become m lattice, wherein m can take following value: m∈{6,12,...,36}.θ (x, y) represents edge orientation map,WhereinIn HMSD, Side vector is turned to m=24 by unification, and step-length is 7.5 °.

In described step b, color quantizing method is as follows: use HSV in the quantization of the color space model of HMSD Color space and the method that RGB unified quantization is 120 colors.H and the R-portion of two of which model are quantified as 0 To 5, S and G part is quantified as 0 to 3, and V and part B are quantized 0 to 4, integrate, and two kinds of models are all It is to produce 120 colors.

Microstructure analysis method in described step c is as follows: in the picture, from top to bottom, from left to right travels through every Individual, when the point (0 °, the consecutive points on 45 ° and 90 ° of directions) of this point and its three neighborhood has one or more points Equal with the value of this point, then this point is marked, and is not marked.When each point of image is through this behaviour After work, just complete the microstructure analysis of image.

It is as follows that micro structure rectangular histogram in described step c describes method: unified image f (x, y) table that will quantify Showing, with f, (x, y)=w represent the value of image；To each some P in image₀=(x₀,y₀), there is f (P₀)=w₀；At 0 °, On 45 ° and 90 ° degree directions, use P_i=(x_i,y_i) represent P₀Three neighborhoods and f (P_i)=w_i, wherein i=1,2,3；w₀With w_iThe secondary N of co-occurrence represent,Represent w₀Occurrence number；On the image quantified from top to bottom, from left to right Travel through each point, add up the point of proximity property associated with it information of this point according to following formula:

Use following formula micro structure rectangular histogram and the micro structure Nogata of color-quantized images respectively to two edge directions Figure merges, H (i)=Max{H_a(i),H_b(i)}；Wherein H_aI () is that under RGB color model, micro structure describes, H_bI () is that under hsv color spatial model, micro structure describes.H (i) be two kinds of micro structures comprehensive after result.

The described micro structure under different colours model describes integrated approach and specifically comprises the following steps that

Obtain the edge direction micro structure rectangular histogram of mixing and the micro structure rectangular histogram of the color-quantized images of mixing, point There is not H_θAnd H_C.Then by H_θAdd at H_CObtain final description below.H_θAnd H_CMerging method as follows Formula.

$H\left(i\right)=\left\{\begin{array}{cc}{H}_C\left(i\right)& i<N\\ \mathrm{\&beta;}\&CenterDot;H_{\mathrm{\&theta;}}(i-N)& i\&GreaterEqual;N\end{array}\right.$

Wherein H is the rectangular histogram description of final HMSD, H_θIt is that edge direction mixing micro structure describes the straight of son Fang Tu, H_CBeing the rectangular histogram of color-quantized images mixing micro structure, β is H_CAnd H_θBetween correlation coefficient, N is H_C Dimension.Because H_CAnd H_θDescribing image different content information, when retrieving image, the degree of contribution is the most different, institute H is revised, in order to improve the feature description ability of H with introducing β.

Beneficial effect

A kind of image search method describing son based on mixing micro structure that the present invention proposes, first with mixing micro-knot Structure describes son (Hybrid MSD, HMSD) and generates the characteristics of image storehouse that image library is corresponding, carries followed by HMSD Take the content characteristic of retrieval image, then utilize similarity measurement criterion to measure, finally by similarity measurement result Sequence, and correspond to image and present to user.

Son is described containing the direction under two kinds of conventional color space model and face owing to the present invention mixing micro structure The micro structure of color, therefore has the strongest resolution capability in color, texture, shape facility and distribution of color information, Reflect the attribute of human visual system to a certain extent.Therefore the image retrieval side of son is described based on mixing micro structure Method makes its image retrieval performance have a distinct increment.

Accompanying drawing explanation

The basic flow sheet of Fig. 1: the inventive method；

Fig. 2: micro structure describes the flow chart that son extracts；

The flow chart of Fig. 3: microstructure analysis；

Detailed description of the invention

In conjunction with embodiment, accompanying drawing, the invention will be further described:

1, to images all in image library, utilizing HMSD to carry out the feature description of picture material, off-line generates figure As the characteristics of image storehouse that storehouse is corresponding；

2, utilize mixing micro structure to describe son and extract the content characteristic of retrieval image；

3, the characteristic vector in the retrieval characteristics of image obtained and characteristics of image storehouse is used L₁Distance carries out phase one by one Like property tolerance.And it is ranked up from small to large according to distance, obtains the front N width result that distance is minimum；

4, it is ranked up from small to large according to distance, obtains the front N width result that distance is minimum.By this N number of result Corresponding image returns.

Mixing micro structure in the present invention, to describe sub-extraction process as follows:

(1) original image of input is respectively converted into the data under RGB and hsv color spatial model, silent The data recognizing reading image are the data under RGB color model, utilize RGB to be converted to hsv color space Computing formula obtain the data of hsv color spatial model.

RGB is converted to the computing formula in hsv color space:

$H=\left\{\begin{array}{cc}\mathrm{\&theta;}& G\&GreaterEqual;B\\ 2\mathrm{\&pi;}-\mathrm{\&theta;}& G<B\end{array}\right\}$

Wherein

$S=1-\frac{\sqrt{3}}{V}min(R,G,B)$

$V=\frac{1}{\sqrt{3}}\&lsqb;R+G+B\&rsqb;$

(2) view data under each color spatial model is carried out respectively edge direction extraction and color quantizing, Obtain color-quantized images C and edge direction image θ；

Described image edge direction extracting method is as follows:

A the gradient in x and the y direction in () color space can be expressed as a (X_h,Y_h,Z_h) and b (X_v,Y_v,Z_v), Wherein X_hRepresent X passage gradient in the horizontal direction.Their norm and dot product can be defined as follows simultaneously:

$\left|a\right|=\sqrt{{\left(X_h\right)}^2+{\left(Y_h\right)}^2+{\left(Z_h\right)}^2}$

$\left|b\right|=\sqrt{{\left(X_v\right)}^2+{\left(Y_v\right)}^2+{\left(Z_v\right)}^2}$

A b=X_hX_v+Y_hY_v+Z_hZ_v

B the angle of () a and b can be expressed as:

$cos(a,b)=\frac{a\&CenterDot;b}{\left|a\right|\&CenterDot;\left|b\right|}$

$\mathrm{\&theta;}=\arccos \left(\frac{a\&CenterDot;b}{\left|a\right|\&CenterDot;\left|b\right|}\right)$

C the edge direction θ unified quantization of calculated each pixel is become m lattice by (), wherein m ∈ 6,12 ..., 36}. θ (x, y) represents edge orientation map,In HMSD, side vector is turned to m=by unification 24, step-length is 7.5 °.

Described color quantizing method is: by hsv color space and method that RGB unified quantization is 120 colors. The H of two of which model and R-portion are quantified as 0 to 5, S and G part is quantified as 0 to 3, V and B portion Dividing and be quantized 0 to 4, so integrating, two kinds of models are all to produce 120 colors.

Particularly for hsv color spatial model: first by H by [0,360) be quantified as [0,179], S and V by [0, 1) be quantified as [0,255], then H is quantified as [0,1 ..., 5], S is quantified as [0,1,2,3], V be quantified as [0,1,2,3, 4].Finally calculated final quantization result by following formula again.

C=H*4*5+S*5+V

For the quantization method broadly similar of RGB color model and HSV, the when of quantization the most for the first time Three passages are all quantized into [0,255]

Traditional color quantizing method be all by hsv color space unified quantization be 8,3,3-dimensional, i.e. H passage Being quantified as 8, S and V is quantified as 3.And the three of RGB color passage constants turn to evenly sized, such as 4、4、4.The characteristics of image that the color quantizing method being found through experiments in the present invention generates more has for image retrieval Benefit.

(3) each color-quantized images and edge direction image carry out microstructure analysis respectively go forward side by side column hisgram Describe；

Described microstructure analysis method is: on image, from top to bottom, from left to right travels through each point, when The value having one or more points and this point in the point of three neighborhoods (0 °, the consecutive points on 45 ° and 90 ° of directions) is equal, So this point is marked, and is not marked.When each point of image is after this operation, just schemed The microstructure image M of picture.Because there being color-quantized images C and edge direction image θ under each color space respectively, Obtain the most altogether four kinds of microstructure images.It is expressed as MC₁、MC₂、Mθ₁、Mθ₂。

Former MSD method is to utilize certain constraints to carry out screening on edge direction image to obtain micro structure.Connect Utilization and just now obtained micro structure as constraints, obtain microstructure image at the enterprising row filter of coloured image quantified. Then on microstructure image, texture and colouring information are added up.Actually take the common factor of two kinds of image informations, Miss substantial amounts of image content information.In HMSD, again micro structure is described son and combined, use also Form efficient combination two kinds of characteristic informations of image of collection, have carried out feature under two kinds of color space model simultaneously and have retouched State.So HMSD is better than original MSD method to the descriptive power of picture material.

Described rectangular histogram describes method: unified by microstructure image f, (x y) represents, the value of image is used (x, y)=w represent f.To each some P in image₀=(x₀,y₀), there is f (P₀)=w₀.At 0 °, 45 ° and 90 ° of degree sides Upwards, P is used_i=(x_i,y_i) represent P₀Three neighborhoods and f (P_i)=w_i, wherein i=1,2,3,.w₀And w_iCo-occurrence time use N represents,Represent w₀Occurrence number.The image quantified travels through each point, and root the most from top to bottom According to following formula statistics it close on property information associated with it.

Wherein, w₀=w_i, { 1,2,3}, α illustrate edge direction Microstructure Information and color-quantized images micro structure to i ∈ The information weight to sign picture material, α carries out approximate evaluation by great many of experiments.

(4) co-occurrence matrix and rectangular histogram is used to carry out the micro structure of two same types under different colours model Describe.The most respectively to { MC₁、MC₂And { M θ₁、Mθ₂Carry out comprehensive description.

The method of feature description is: described in the rectangular histogram due to four micro structures, grain direction Microstructure Information is straight Side's figure is 24 dimensions, and the rectangular histogram describing color-quantized images Microstructure Information is 120 dimensions.If it is straight by these four If the directly combination of side's figure, total rectangular histogram just has 288 dimensions.Dimension is excessive when causing feature extraction and similarity system design Between and resource cost be unacceptable.It is thus desirable to a kind of effective method reduces histogrammic dimension.Use following formula Respectively the micro structure rectangular histogram of two edge directions and the micro structure rectangular histogram of color-quantized images are merged,

H (i)=Max{H₁(i),H₂(i)}

Wherein H₁I () is that under RGB color model, micro structure describes (MC₁、Mθ₁), H₂I () is HSV face Under color space model, micro structure describes (MC₂、Mθ₂).H (i) be two kinds of micro structures comprehensive after result, mixed Edge direction micro structure rectangular histogram and the micro structure rectangular histogram of color-quantized images of mixing, have H respectively_θAnd H_C。

(5) the color-quantized images micro structure rectangular histogram of mixing and the edge direction micro structure rectangular histogram of mixing are entered Row comprehensive description.Description method is:

There is H respectively_θAnd H_C.Then by H_θAdd at H_CObtain final description below.H_θAnd H_CConjunction And method such as following formula.

$H\left(i\right)=\left\{\begin{array}{cc}{H}_C\left(i\right)& i<N\\ \mathrm{\&beta;}\&CenterDot;H_{\mathrm{\&theta;}}(i-N)& i\&GreaterEqual;N\end{array}\right.$

Wherein H is the rectangular histogram description of final HMSD, H_θIt is that edge direction mixing micro structure describes the straight of son Fang Tu, H_CBeing the rectangular histogram of color-quantized images mixing micro structure, β is H_CAnd H_θBetween correlation coefficient, N is H_CDimension.Because H_CAnd H_θQuantify dimension and the feature difference from the effect in retrieval, so introducing β revises H, plays the effect of the feature description ability improving H.

The present invention uses L₁Distance is as similarity measurement, and formula is as follows:

$L_1(A,B)=\underset{i=1}{\overset{n}{\&Sigma;}}\left|a_i-b_i\right|$

Wherein A, B are two width image characteristic of correspondence vectors, a_i、b_iRepresentative feature component.

Claims (5)

1. the image search method describing son based on mixing micro structure, it is characterised in that step is as follows:

Step 1: to images all in image library, utilizes mixing micro structure to describe son and carries out the feature description of picture material, Generate corresponding characteristics of image storehouse；

Step 2: utilize mixing micro structure to describe son and extract the content characteristic of retrieval image；

Step 3: the characteristic vector in retrieval characteristics of image and characteristics of image storehouse is carried out similarity measurement；

Step 4: returned by image corresponding for the top n result of similarity measurement, similarity quantization method uses city Block distance；

It is as follows that the mixing micro structure of described step 1 and step 2 describes sub-step:

Step a: input picture is respectively converted into the data under RGB and hsv color spatial model；

Step b: the view data under each color spatial model is carried out respectively edge direction extraction and color quantizing, Obtain color-quantized images and edge direction image；

Step c: respectively each color-quantized images and edge direction image are carried out microstructure analysis and goes forward side by side column hisgram Describe；

Step d: use co-occurrence matrix and rectangular histogram to carry out the micro structure of two same types under different colours model Describe；

Step e: the color-quantized images micro structure rectangular histogram of mixing and the edge direction micro structure rectangular histogram of mixing are entered Row is comprehensive；

Microstructure analysis method in described step c is as follows:

In the picture, from top to bottom, from left to right travel through each point, when in the parameter value of this point with the point of its three neighborhood The corresponding parametric values of any point is equal, then this point is marked, and is not marked；When image each some warp After crossing this operation, just complete the microstructure analysis of image；Described three neighborhoods refer to: for any pixel point, With this point as the center of circle, with it with first, right side line as radius, rotate 0 °, 45 ° and 90 ° the most respectively Each consecutive points on three directions, these three point collectively forms three neighborhoods.

The image search method of son is described based on mixing micro structure the most according to claim 1, it is characterised in that: described Edge direction extraction step in step b is as follows:

Step (a): use Sobel operator to obtain each passage of image in x and y direction in three-dimensional color space Direction gradient；

Step (b): according to formulaObtain formula A b=X_hX_v+Y_hY_v+Z_hZ_v, then according to angle formulae between vectorCalculate image border Directional image；

Wherein, a (X_h,Y_h,Z_h) and b (X_v,Y_v,Z_v) represent gradient both horizontally and vertically, X respectively_hRepresent X passage Gradient in the horizontal direction, X_vRepresent the X passage gradient in vertical direction；Y_h、Y_vAnd Z_h、Z_vIt is that Y, Z are logical respectively Road is in the gradient of both direction；

Step (c): the edge direction θ unified quantization of each pixel is become m lattice, wherein m can take following value: m∈{6,12,...,36}；θ (x, y) represents edge orientation map,Wherein

The image search method of son is described based on mixing micro structure the most according to claim 2, it is characterised in that: described When in step b, edge direction unified quantization to each pixel becomes m lattice, m=24, step-length is 7.5 °.

The image search method of son is described based on mixing micro structure the most according to claim 1, it is characterised in that: described In step b, color quantizing method is as follows: use hsv color space in the quantization of the color space model of HMSD With the method that RGB unified quantization is 120 colors；The H of two of which model and R-portion are quantified as 0 to 5, S Being quantified as 0 to 3 with G part, V and part B are quantized 0 to 4, integrate, and two kinds of models are all to produce Raw 120 colors.

The image search method of son is described based on mixing micro structure the most according to claim 1, it is characterised in that: described It is as follows that micro structure rectangular histogram in step c describes method: unified image f (x, y) expression, the figure that will quantify With f, (x, y)=w represent the value of picture；To each some P in image₀=(x₀,y₀), there is f (P₀)=w₀；At 0 °, 45 ° and 90 ° degree directions on, use P_i=(x_i,y_i) represent P₀Three neighborhoods and f (P_i)=w_i, wherein i=1,2,3；w₀ And w_iThe secondary N of co-occurrence represent, N represents w₀Occurrence number；Quantify image on from top to bottom, from a left side To each point of right traversal, add up the point of proximity property associated with it information of this point according to following formula:

Following formula is used respectively the micro structure rectangular histogram of two edge directions and the micro structure rectangular histogram of color-quantized images to be entered Row merges, H (i)=Max{H_a(i),H_b(i)}；Wherein H_aI () is that under RGB color model, micro structure describes, H_bI () is that under hsv color spatial model, micro structure describes；H (i) be two kinds of micro structures comprehensive after result.