CN105701174A - Appearance design patent retrieving method based on dynamic texton - Google Patents

Abstract

The present invention relates to an appearance design patent retrieving method based on a dynamic texton. The method comprises the following steps: step 1) extracting a texton; and step 2) extracting an LBP feature based on statistical analysis from the texton, and performing normalization and similarity measurement on an extracted feature vector, and returning a retrieval result. According to the method provided by the present invention, on the basis of an appearance design patent and targeted at the characteristic that different image textons may have unequal sizes, an image retrieving algorithm of dynamcially extracting a texton according to image content and extracting a feature from the texton is provided. When retrieving images with a repeated texton, the retrieval effect of the method provided by the present invention is superior to the existing retrieving method based on a whole image, and images with the same texton but different arrangement structures can be searched out; and the method can be used for image texture structure similarity determination, has excellent retrieval effects and can well meet the requirements of actual application.

Description

A kind of design patent retrieval method based on dynamic texture primitive

Technical field

The invention belongs to design patent retrieval technical field, be specifically related to a kind of design patent retrieval method based on dynamic texture primitive。

Background technology

Along with the quickening of the development of international-style and trade integration process, design patent occupies increasingly consequence in industrial property, and design patent application quantity quickly increases。Design patent, using image as protection object, mainly protects the foreground object in patent image。For design patent, the contrast that patent examination personnel will submit to the patent image applied in the appearance patent image applied for and data base to carry out shape, pattern, color, it is judged that whether it has certain similarity。By the similarity criterion of design patent image it can be seen that for containing identical texture primitive, but because arrangement mode or number are different and present two width appearance patent image of different pattern content, identical appearance design should be judged to。Therefore, when carrying out design patent retrieval, identical texture primitive can be retrieved but the image of different arranged distribution or number is particularly important。Owing to texture primitive has certain diversity in spatial arrangements rule, this type of image can not be all retrieved by the image search method being generally basede on content。The method carrying out retrieving based on texture primitive can avoid image texture primitive arrangement mode or the impact comprising number, and retrieval has the image of identical texture primitive, thus that avoids design patent repeats application, design patent retrieval has certain meaning。

The current conventional search method based on texture primitive is to divide an image into equal-sized image block, and using image block as texture primitive, and then extract primitive feature realization retrieval。Local description can also, as texture primitive, be used for describing texture, and can be used for the local description of this framework has SIFT, SURF etc.。The content comprised due to different classes of image is different, and the size of texture primitive also differs, and therefore arranging texture primitive is that fixed size can not correctly reflect the real texture information of image。Based on the search method of texture primitive spatial distribution characteristic, space distribution information is introduced textural characteristics descriptor, improve the recall precision of image, but the image that spatial distribution is different for texture is identical, this search method can not obtain good retrieval effectiveness。

Although the existing search method based on texture primitive proposes the concept of texture primitive, but can not effectively extract the texture primitive in image。

Summary of the invention

For above-mentioned problems of the prior art, it is an object of the invention to provide a kind of design patent retrieval method based on dynamic texture primitive avoiding the occurrence of above-mentioned technological deficiency。

In order to realize foregoing invention purpose, the technical solution used in the present invention is as follows:

A kind of design patent retrieval method based on dynamic texture primitive, comprises the following steps:

Step 1): extract texture primitive；

Step 2): texture primitive is extracted the LBP feature that Corpus--based Method is analyzed, and the characteristic vector extracted is normalized and similarity measurement, return retrieval result。

Further, described step 1) particularly as follows:

Step A: the image of texture primitive to be extracted is carried out the pretreatment operation such as gray processing, denoising, normalization；

Step B: image being done LBP conversion, extracts initial texture primitive, such as formula (3), formula (4), formula (5), shown in formula (6)；

Wherein: LBP converts as shown in formula (1) and formula (2):

If original image matrix I=(p_ij)_h×w, p_ijFor at (i, j) pixel value at place。Image array after LBP converts is I_LBP=(P '_ij)_h×w, p '_ijFor p_ijValue corresponding after doing LBP conversion。LBP conversion process such as formula (1):

$p_{ij}^{\′}=\underset{l=0}{\overset{2^N-1}{\Σ}}{2}^l\×S(g_l-g_c)---\left(1\right)$

$S(g_l-g_c)=\left\{\begin{array}{c}1,g_l\&GreaterEqual;g_c\\ 0,g_l<g_c\end{array}\right.---\left(2\right)$

Wherein, g_cCentered by some pixel value, i.e. p_ij, g_lFor neighborhood territory pixel value, N is the radius of neighbourhood。

$I_i=\left\{\begin{array}{c}True,abs(L_i-Lmin)\&le;Lratio\&times;(L\max -Lmin)\\ False,abs(L_i-Lmin)>Lratio\&times;(L\max -Lmin)\end{array}\right.---\left(3\right)$

H=min (I_i-I_i+1)(4)

$R_i=\left\{\begin{array}{c}True,abs(R_i-Rmin)\&le;Rratio\&times;(R\max -Rmin)\\ False,abs(R_i-Rmin)>Rratio\&times;(R\max -Rmin)\end{array}\right.---\left(5\right)$

W=min (R_i-R_j)(6)

Wherein, L_i, projection value that i ∈ [0, h-1] is horizontal direction, computing formula such as formula (7) and shown in formula (8)；R_j, projection value that j ∈ [0, w-1] is vertical direction, computing formula such as formula (9) and shown in formula (10)；

$L_i=\underset{j=0}{\overset{w-1}{\&Sigma;}}{L}_j^{\&prime;},0\&le;i\&le;h-1---\left(7\right)$

$L_j^{\&prime;}=\left\{\begin{array}{c}1,p_{ij}^{\&prime;}=Num\\ 0,p_{ij}^{\&prime;}\&NotEqual;Num\end{array}\right.---\left(8\right)$

$R_j=\underset{i=0}{\overset{h-1}{\&Sigma;}}{R}_i^{\&prime;},0\&le;j\&le;w-1---\left(9\right)$

$R_i^{\&prime;}=\left\{\begin{array}{c}1,p_{ij}^{\&prime;}=Num\\ 0,p_{ij}^{\&prime;}\&NotEqual;Num\end{array}\right.---\left(10\right)$

Step C: draw respectively to the right and downwards take onesize with initial texture primitive and not superimposed images block as sliding window；

Step D: judge initial texture primitive and draw whether take image block equal, as shown in formula (11)；If it is equal with initial texture primitive to only exist a sliding window, go to step E；If two sliding windows are all unequal with initial texture primitive, expand initial texture primitive, go to step C；If two sliding windows are all equal with initial texture primitive, go to step E；

$isBE(block1,block2)=\left\{\begin{array}{cc}1,& per\&GreaterEqual;threshold2\\ 0,& per<threshold2\end{array}\right.---\left(11\right)$

Wherein,

$per=\frac{\underset{i=0}{\overset{h-1}{\&Sigma;}}\underset{j=0}{\overset{w-1}{\&Sigma;}}isEqual\&lsqb;p_1(i,j),p_2(i,j)\&rsqb;}{H\&times;W}---\left(12\right)$

$isEqual\&lsqb;p_{i,j},p_{i,j}^{\&prime;}\&rsqb;=\left\{\begin{array}{c}1,abs\&lsqb;p_{i,j}-p_{i,j}^{\&prime;}\&rsqb;\&le;threshold\\ 0,abs\&lsqb;p_{i,j}-p_{i,j}^{\&prime;}\&rsqb;>threshold\end{array}\right.---\left(13\right)$

Wherein, p_{I, j}, p '_{I, j}Pixel value for different images block corresponding pixel points。

Step E: if the sliding window that there is certain direction is unequal with initial texture primitive, obtains new sliding window at direction slip certain distance in allowed band, go to step D；

Step F: using current initial texture primitive as image texture primitive。

Further, described step 2) in extract LBP feature particularly as follows:

Use " equivalent formulations " that the schema category of LBP operator is carried out dimensionality reduction, as shown in formula (14):

${\mathrm{LBP}}_{P,R}^{riu2}=\left\{\begin{array}{cc}\underset{i=0}{\overset{P-1}{\&Sigma;}}S(g_i-g_c)& U\left({\mathrm{LBP}}_{P,R}\right)\&le;2\\ P+1& otherwise\end{array}\right.---\left(14\right)$

Wherein, $U\left({\mathrm{LBP}}_{P,R}\right)=|S(g_{P-1}-g_c)-S(g_0-g_c)|+\underset{i=0}{\overset{P-1}{\&Sigma;}}|S(g_i-g_c)-S(g_{i-1}-g_c)|,$ If x >=0, S (x) takes 1, and otherwise S (x) takes 0, g_cRepresent the gray value of center pixel, g in neighborhood_iRepresenting the gray value of each pixel in the border circular areas that radius is R, wherein the span of i is 0 to P-1。

Design patent retrieval method based on dynamic texture primitive provided by the invention, based on design patent, for different images texture primitive fixed equal feature not of uniform size, it is proposed to a kind of according to picture material dynamic extraction texture primitive and the image retrieval algorithm that texture primitive is extracted feature。When retrieval has the image of repetition texture primitive, the retrieval effectiveness of the present invention is better than the existing search method based on entire image, can retrieve and there is identical texture primitive, but the image that arrangement architecture is different, can be used for image texture structural similarity to judge, there is good retrieval effectiveness, it is possible to meet the needs of practical application well。

Accompanying drawing explanation

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

Fig. 2 is the LBP value that different images block is corresponding；

Fig. 3 is identical texture primitive difference separating capacity；

Fig. 4 is texture primitive skew；

Fig. 5 is for dividing texture primitive block；

Fig. 6 is threshold value is that texture primitive when 0 extracts result；

When Fig. 7 is fixed threshold, texture primitive extracts result；

Fig. 8 is the relation between threshold value value and texture primitive accuracy；

Fig. 9 is that texture primitive offsets；

Figure 10 is that texture primitive Heterogeneous Permutation generates new images；

Figure 11 is identical texture primitive different images content。

Detailed description of the invention

In order to make the purpose of the present invention, technical scheme and advantage clearly understand, below in conjunction with the drawings and specific embodiments, the present invention will be further described。Should be appreciated that specific embodiment described herein is only in order to explain the present invention, is not intended to limit the present invention。

As it is shown in figure 1, a kind of design patent retrieval method based on dynamic texture primitive, comprise the following steps:

Step 1): extract texture primitive；

Texture primitive is texture structure basic in image, describes the Local textural feature of image, occurs by certain queueing discipline cycle in the picture。Texture primitive is relied on to contribute to by different types of object areas separately。Can obtaining texture primitive by the texture information of analysis image and have the following characteristics that (1) nonuniqueness, the texture primitive of piece image can have multiple。Starting point is different, and the texture primitive presented is different；(2) same image texture primitive size is fixed, and the texture primitive of piece image can not uniquely, but size is fixed；(3) sliceable property, though image texture primitive is not unique, but spliced image is identical。Property repeatable, sliceable and the unique feature of size because of texture primitive, it is determined that after texture primitive, primitive should not overlapping with neighborhood and formed objects image block essentially equal。Therefore the extraction of texture primitive can be converted to the minimum multiimage's block of extraction。

Step 2): texture primitive is extracted the LBP feature that Corpus--based Method is analyzed, and the characteristic vector extracted is normalized and similarity measurement, return retrieval result。

Wherein, described step 1) particularly as follows:

Step A: the image of texture primitive to be extracted is carried out the pretreatment operation such as gray processing, denoising, normalization；

Step B: image is done LBP conversion, extracts initial texture primitive；

LBP is a kind of operator describing image local textural characteristics, embodies the gray-value variation of image pixel vertex neighborhood。The grey scale change of LBP value this vertex neighborhood pixel of more big expression is more violent, corresponds to high-frequency information, describes the contour structure of image；The grey scale change being worth this vertex neighborhood point of more little expression is more slow, corresponds to low-frequency information, describes the entire content of image。According to the LBP operator descriptive power to image texture structure, and texture primitive has periodically, and corresponding LBP value also should present period profile, and the cycle of repetition is height value corresponding to initial texture primitive and width value。If the minima of initial texture primitive size is H_initial×W_initial。

If original image matrix I=(p_ij)_h×w, p_ijFor at (i, j) pixel value at place。Image array after LBP converts is I_LBP=(p '_ij)_h×w, p '_ijFor p_ijValue corresponding after doing LBP conversion。LBP converts as shown in formula (1) and formula (2):

$p_{ij}^{\&prime;}=\underset{l=0}{\overset{2^N-1}{\&Sigma;}}{2}^l\&times;S(g_1-g_c)---\left(1\right)$

$S(g_l-g_c)=\left\{\begin{array}{c}1,g_l\&GreaterEqual;g_c\\ 0,g_l<g_c\end{array}\right.---\left(2\right)$

Wherein, g_cCentered by some pixel value, i.e. p_ij, g_lFor neighborhood territory pixel value, N is the radius of neighbourhood。

Image is after LBP converts, and the point that LBP takes a certain fixed value is added up, and both horizontally and vertically projects respectively。Wherein, the projection value of horizontal direction is L_i, i ∈ [0, h-1], the height value of repetition period correspondence initial texture primitive in the horizontal direction。Shown in computing formula such as formula (3) and formula (4)；

$L_i=\underset{j=0}{\overset{w-1}{\&Sigma;}}{L}_j^{\&prime;},0\&le;i\&le;h-1---\left(3\right)$

$L_j^{\&prime;}=\left\{\begin{array}{c}1,p_{ij}^{\&prime;}=Num\\ 0,p_{ij}^{\&prime;}\&NotEqual;Num\end{array}\right.---\left(4\right)$

The projection value of vertical direction is R_j, j ∈ [0, w-1], repetition period correspondence initial texture element width value in the vertical direction, computing formula such as formula (5) and shown in formula (6)；

$R_j=\underset{i=0}{\overset{h-1}{\&Sigma;}}{R}_i^{\&prime;},0\&le;j\&le;w-1---\left(5\right)$

$R_i^{\&prime;}=\left\{\begin{array}{c}1,p_{ij}^{\&prime;}=Num\\ 0,p_{ij}^{\&prime;}\&NotEqual;Num\end{array}\right.---\left(6\right)$

Projection when LBP value takes a certain fixed value is added up in the horizontal direction after LBP conversion。Because picture material has certain repeatability, the distribution of LBP value also presents periodically, and the repetition period is the height value of texture primitive。Distribution according to LBP value, intercepting the minimum length value between adjacent minimum point is the repetition period。Wherein shown in the decision method of minimum point such as formula (7) and formula (8):

$I_i=\left\{\begin{array}{c}True,abs(L_i-Lmin)\&le;Lratio\&times;(L\max -Lmin)\\ False,abs(L_i-Lmin)>Lratio\&times;(L\max -Lmin)\end{array}\right.---\left(7\right)$

H=min (I_i-I_i+1)(8)

Wherein, L_iTaking projection during a certain value for each row LBP value in the horizontal direction, Lmin is projection minima, and Lmax is projection maximum, and Lratio is coefficient, and the point meeting formula (7) is marked as minimum point, is designated as I_i, wherein i ∈ [1, n], n is minimum point sum。As shown in formula (8), I_i, I_i+1It is two adjacent minimum point, i ∈ [1 ..., n-1]。All adjacent minimum point are calculated difference between two, takes the length that difference minima is the cycle, the i.e. height value of initial texture primitive, and H >=H_initial。

In like manner, projection when statistics each column LBP value takes a certain value in the vertical direction, the maxima and minima in acquisition projection value, the point meeting formula (9) is labeled as minimum point, is designated as R_i。The width value that minima is initial texture primitive of the distance between all adjacent minimum point, and W >=W_initial, as shown in formula (10)。

$R_i=\left\{\begin{array}{c}True,abs(R_i-Rmin)\&le;Rratio\&times;(R\max -Rmin)\\ False,abs(R_i-Rmin)>Rratio\&times;(R\max -Rmin)\end{array}\right.---\left(9\right)$

W=min (R_i-R_j)(10)

Step C: draw respectively to the right and downwards take onesize with initial texture primitive and not superimposed images block as sliding window；

Though LBP value can reflect the texture information of image local, the intensity of variation of image pixel value can not be reflected。As in figure 2 it is shown, (a), (b) is the image block that two texture structures are incomplete same。A the pixel value of () image block is comparatively average, the pixel value excursion of (b) image block is relatively big, but after image block does LBP conversion, two image blocks obtain identical LBP value。Image carries out LBP conversion, and dissimilar image block is likely to be obtained identical LBP value, thus causing that the texture primitive extracted is imperfect。Therefore, the present invention using by the texture primitive of LBP change detection as initial texture primitive blocks, extract complete texture primitive further by the method for individual element point contrast images block。

Definition 1: image separating capacity refers to the degree distinguished of foreground object and background color。If display foreground object color is more close with background color, it is easy to obscuring foreground object and background color, now the separating capacity of image is more weak, otherwise, if foreground color and background color difference are relatively big, the separating capacity of image is stronger。As it is shown on figure 3, figure (a) and (b) for texture primitive is identical but two width images that background color is different。The foreground object color of figure (a) is close with background colour, and the separating capacity of image is more weak, schemes (b) foreground object color and background colour obvious difference, and separating capacity is stronger。

Definition 2: texture primitive skew refers to that texture primitive is not in same level or vertical direction order arrangement, and down or toward dextroposition skew certain distance, move two kinds of situations including lower skew and right avertence。As shown in Figure 4, (a) is artwork, and (b) is the texture primitive that figure (a) is corresponding, and (c) splices the image obtained for texture primitive。By figure (c) it can be seen that when texture primitive splice, and manage the image that primitive splicing obtains。By figure (c) it can be seen that when texture primitive is spliced, be not sequential concatenation, but offset to the right certain distance splicing and obtain original image。Therefore, the texture primitive scheming (a) there occurs that right avertence is moved。By analyzing the skew of texture primitive it can be seen that lower skew and right avertence are moved and will not be occurred simultaneously。Draw downwards and take image block when judging similarity, need to judge that right avertence is moved, the situation of lower skew will not occur；In like manner, draw to the right and take image block when judging similarity, only need to judge lower skew, the situation that right avertence is moved will not occur。

Step D: judge initial texture primitive and draw whether take image block equal。If it is equal with initial texture primitive to only exist a sliding window, go to step E；If two sliding windows are all unequal with initial texture primitive, expand initial texture primitive, go to step C；If two sliding windows are all equal with initial texture primitive, go to step E；

After determining initial texture primitive, turning left by order shown in Fig. 5 and down draw and take onesize image block, compare with initial texture primitive individual element point respectively, I_initial is initial texture primitive, I_Left draws the image block taken to the left, and I_Right draws downwards the image block taken。Pixel compares shown in formula such as formula (11)。I_initial is put pixel-by-pixel with I_Left, I_Right respectively and compares, if difference is equal less than threshold determination two pixel, otherwise not etc.。If all pixels of whole image block are all equal, then two image blocks are identical, otherwise differ。

$isEqual\&lsqb;p_{i,j},p_{i,j}^{\&prime;}\&rsqb;=\left\{\begin{array}{c}1,abs\&lsqb;p_{i,j}-p_{i,j}^{\&prime;}\&rsqb;\&le;threshold\\ 0,abs\&lsqb;p_{i,j}-p_{i,j}^{\&prime;}\&rsqb;>threshold\end{array}\right.---\left(11\right)$

Wherein, p_{I, j}, p '_{I, j}Pixel value for different images block corresponding pixel points。

Fig. 6 is threshold when being 0, and texture primitive extracts result, and in figure (a), (b), the texture of (c) has obvious repeatability, but the texture primitive extracted is image itself。By the impact of picture noise, visually identical image, the pixel value of corresponding point might not be essentially equal。If only assert when pixel value is essentially equal, two pixels are identical, and because influence of noise cannot be completely eliminated, the texture primitive of extraction can only be image itself, extract texture primitive and just lose meaning。Therefore should allow there is certain error between pixel value, realize image block phase by step-up error threshold value

Fig. 7 is that threshold takes texture primitive extraction result during fixed value。Image (a), (d) is original image, (b), (c) and (e), (f), and (g) is the corresponding texture primitive extracted。It can be seen that some image can correctly extract primitive, some image then can not。The separating capacity of different images is different, if error threshold is set to fixed value, the picture stronger for separating capacity can correctly extract texture primitive, for the picture that separating capacity is more weak, easily think background colour by mistake prospect, cause that extraction texture primitive is imperfect or extracts repetition。Therefore, the error threshold of image should be dynamically adjusted according to image self information。

When threshold value is bigger the tolerance of picture noise is relatively good, when threshold value is less, the separating capacity of background and foreground object is better。Variance is used for the departure degree measuring stochastic variable with average, and the present invention judges the separating capacity of image by mean variance。Take average and variance respectively avg and σ of image and average difference, computing formula such as formula (12) and shown in formula (13)。σ is more big, it was shown that in image, the span of pixel is more wide, and namely the separating capacity of image is more strong。Otherwise separating capacity is more weak。

$avg=\frac{1}{h\&times;w}\underset{i=1}{\overset{h}{\&Sigma;}}\underset{j=1}{\overset{w}{\&Sigma;}}abs\left(\frac{1}{h\&times;w}\underset{i=1}{\overset{h}{\&Sigma;}}\underset{j=1}{\overset{w}{\&Sigma;}}p\right(i,j)-p(i,j\left)\right)---\left(12\right)$

$\mathrm{\&sigma;}=\sqrt{\frac{1}{h\&times;W}\underset{i=1}{\overset{h}{\&Sigma;}}\underset{j=1}{\overset{w}{\&Sigma;}}{\left(abs\right(\frac{1}{h\&times;w}\underset{i=1}{\overset{h}{\&Sigma;}}\underset{j=1}{\overset{w}{\&Sigma;}}p\left(i,j\right)-p\left(i,j\right))-avg)}^2}---\left(13\right)$

$threshold=\left\{\begin{array}{cc}ratio1\&times;avg& \mathrm{\&sigma;}\&le;thres\\ ratio2\&times;avg& otherwise\end{array}\right.---\left(14\right)$

In above formula, avg is that the difference with image average is average, and σ is the departure degree with average, and threshold is error threshold, and thres is the threshold value of departure degree, and ratio1 and ratio2 is different coefficient。When variance is less than or equal to thres, the separating capacity of image is more weak, and error threshold coefficient is ratio1, otherwise is ratio2。

Judge that arranging threshold value when pixel value is equal suppresses effect of noise to a certain extent, but can not be completely eliminated。Therefore, for reducing noise, texture primitive is extracted the impact of result, threshold value threshold2 is set when finally judging image block similarity, represent that identical pixels is counted out and account for whole image block pixel number purpose percentage ratio。Threshold2 is more little, and the tolerance power of noise is more big。When judging the similarity of image block, if identical pixels is counted out, percentage is be more than or equal to the threshold value threshold2 arranged, then two image blocks are equal, otherwise not etc.。The image that separating capacity is more weak, foreground and background color is comparatively similar, if threshold value is too small, it is easy to think part background color by mistake foreground, thus causing that the texture primitive extracted is imperfect, if threshold value is excessive, the fault-tolerant ability of noise can be reduced。Therefore, it is necessary to select suitable threshold value。

Fig. 8 is threshold2 value and the relation extracted between texture primitive accuracy。As seen from the figure, along with the increase of threshold2, extract texture primitive within the specific limits and correctly take the lead in declining after rising。Threshold2 arranges too small, it is easy to cause that the texture primitive of the more weak image zooming-out of separating capacity is imperfect, thus affecting accuracy。If what arrange is excessive, the fault-tolerant ability of noise can be reduced, cause that the texture primitive extracted is not minimum texture structure, texture primitive accuracy is had certain impact equally。Therefore, the image that separating capacity is stronger, what should suitably be arranged by threshold2 is big, and the image that separating capacity is more weak, then what should arrange is relatively small。Judge shown in the formula such as formula (15) and (16) that image block is equal:

$isBE(block1,block2)=\left\{\begin{array}{cc}1,& per\&GreaterEqual;threshold2\\ 0,& per<threshold2\end{array}\right.---\left(15\right)$

$per=\frac{\underset{i=0}{\overset{h-1}{\&Sigma;}}\underset{j=0}{\overset{w-1}{\&Sigma;}}isEqual\&lsqb;p_1(i,j),p_2(i,j)\&rsqb;}{H\&times;W}---\left(16\right)$

The texture of image might not arrange according to rule ordering, it is possible that offset in extracting texture primitive process。If the image block sliding window that order is chosen is unequal with initial texture primitive, do not indicate that initial texture primitive is unsatisfactory for the condition of texture primitive, and sliding window should waited, do not indicate that initial texture primitive is unsatisfactory for the condition of texture primitive, and should continue to judge at direction, place slip certain distance by sliding window。Right avertence is moved and lower skew only there will be either or both and all occurs without, and the distance of skew should be less than the size of now initial texture primitive, and namely right side-play amount should be less than the width of initial texture primitive, and lower side-play amount then should be less than the height of initial texture primitive。As it is shown in figure 9, (a) is original image, (b) is the texture primitive extracted, and figure (c) is the image that texture primitive splicing is obtained, and can be seen that texture primitive there occurs that right avertence is moved by scheming (c)。

Step E: if the sliding window that there is certain direction is unequal with initial texture primitive, obtains new sliding window at direction slip certain distance in allowed band, go to step D；

Step F: using current initial texture primitive as image texture primitive。

Textural characteristics is one of key character of image, and essence portrays the neighborhood gray space regularity of distribution of pixel, and the spatial information in image is done quantitative description。Compared to other characteristics of image, textural characteristics can better take into account image macroscopic property and two aspects of microstructure。

Local binary patterns (LocalBinaryPattern) is a kind of operator for describing image local textural characteristics, and the textural characteristics being based on N*N statistics describes method。It is low that this operator has computation complexity, unordered training study, illumination invariant and be prone to the advantages such as Project Realization。The starting point that starting point is texture primitive of default image in the process of the inventive method extraction texture primitive, the texture primitive of the image zooming-out of thus like texture primitive is identical, but spatial distribution is likely to difference。LBP descriptor is more weak to the descriptive power of image tiny model spatial relationship, is left out LBP value information in spatial distribution, therefore texture primitive is extracted LBP feature and can avoid the problem that spatial distribution is different。

Original LBP operator definitions is N*N window, cover only region only small within the scope of radii fixus, and the robustness of texture localized variation is poor, it is impossible to extract the key feature of large scale texture。In order to overcome these shortcomings, LBP operator has been improved by T.Ojala etc., and the LBP operator after improvement can expand to any neighborhood, and circle shaped neighborhood region instead of original square neighborhood, it is allowed to have any number of sampled point in circle shaped neighborhood region。LBP operator after improvement describes circle shaped neighborhood region with (P, R), and wherein P represents number of pixels in neighborhood, and R represents the radius of neighbourhood。

One LBP operator can produce much different binary modes, and along with the increase of sampling number in neighborhood collection, the kind of binary mode increases accordingly。Too much binary mode is not only unfavorable to the statement of texture, identification, classification etc., and engineer applied causes very big difficulty equally。In order to solve the problem that binary mode is too much brought, T.Ojala etc. proposes the one " equivalent formulations " schema category to LBP operator and carries out dimensionality reduction。If the coding corresponding to a local binary pattern only comprises at most once from 0 to 1 with once from the saltus step of 1 to 0, then claiming to be encoded to equivalent formulations class corresponding to this local binary pattern, pattern T.Ojala now is referred to as Uniform pattern。Such as: 00000000,00111000 and 11100001 is all referred to as Uniform pattern。This LBP mode flag it isWherein, described step 2) in extract LBP feature particularly as follows:

Use " equivalent formulations " that the schema category of LBP operator is carried out dimensionality reduction, as shown in formula (17):

${\mathrm{LBP}}_{P,R}^{riu2}=\left\{\begin{array}{cc}\underset{i=0}{\overset{P-1}{\&Sigma;}}S(g_i-g_c)& U\left({\mathrm{LBP}}_{P,R}\right)\&le;2\\ P+1& otherwise\end{array}\right.---\left(17\right)$

Wherein, $U\left({\mathrm{LBP}}_{P,R}\right)=|S(g_{P-1}-g_c)-S(g_0-g_c)|+\underset{i=0}{\overset{P-1}{\&Sigma;}}|S(g_i-g_c)-S(g_{i-1}-g_c)|,$ If x >=0, S (x) takes 1, and otherwise S (x) takes 0, g_cRepresent the gray value of center pixel, g in neighborhood_iRepresenting the gray value of each pixel in the border circular areas that radius is R, wherein the span of i is 0 to P-1。

Wherein for described step 2) in, when repeatedly contrast experiment takes 3 as α, effect is best。

Identical texture primitive, different arrangement modes can produce different picture materials。Present invention experiment is chosen some width texture primitives in image library and is had the image zooming-out texture primitive of repeatability, then each texture primitive Heterogeneous Permutation is generated 3 width different images。As shown in Figure 10, (a) is the texture primitive extracted, and (b)-(d) adopts different side-play amounts that texture primitive carries out the image that dislocation splicing obtains。The texture primitive of image (b)-(d) is all figure (a), but be there is the image of repeatability by different modes structure, after extracting texture primitive, it is spliced, each image can obtain, by splicing, the image that 3 width are different, therefore obtains the image library with 945 images。

In the inventive method, the minima arranging initial texture element height and width is respectively as follows: H_initial=32, W_initial=32。In image, noise point belongs to high fdrequency component, and the more little impact by noise point of LBP value is more little, therefore chooses the point that LBP value is 0 when determining the size of initial texture primitive and adds up, it is to avoid effect of noise。

Judge that the extraction on texture primitive of choosing of the threshold value threshold2 similar to judging image block for threshold value threshold that pixel is equal has certain impact, therefore need to arrange suitable threshold value in the process extracting texture primitive。It is 0.8 that the present invention arranges ratio1 and ratio2 difference 1 and 1.4, threshold2。Image is done normalization pretreatment by the inventive method before extracting texture primitive, and unification is set to 256*256。But the texture primitive number that identical texture primitive and formed objects image comprise is not necessarily identical, thus the texture primitive size extracted is unequal。As shown in figure 11, image (a), (b) is the image with identical texture primitive being normalized to 256*256, and texture primitive is all figure (b), and spliced image is all figure (c)。But figure (a) comprises 2 texture primitives, and figure (b) only comprises a texture primitive。Therefore, although image (a), the texture primitive of (c) is identical, but the texture primitive size extracted is unequal。For avoiding image to differ in size, retrieval result being impacted, the present invention extracts the LBP feature that Corpus--based Method is analyzed, and the characteristic vector extracted is normalized。

The present invention adopts following Performance Evaluating Indexes:

Definition 3: be identical image with the image definition of color with texture primitive；

Definition 4: be similar image with texture primitive but the image definition of different colours；

(1) identical image pertinency factor STSC

STSC (SameTextureSameColor) represents have identical texture primitive and the image pertinency factor that color is identical with the image that is retrieved。Namely before, N opens in image with the image that is retrieved with the texture image percentage with color。

$STSC=\frac{SameNum}{N}---\left(18\right)$

Wherein, SameNum represents picture number identical with the image that is retrieved in the front N width image of return result, and N is total number of images order identical with the image that is retrieved in whole image library。The STSC measure algorithm pertinency factor to identical image, STSC value is more big, and the image representing identical with the image that is retrieved is more forward, and retrieval effectiveness is more good。Otherwise, retrieval effectiveness is more poor。

(2) similar image pertinency factor STDC

STDC (SameTextureDifferentColor) represents the image retrieval rate with the image similarity that is retrieved。Namely before, M opens in image and the image percentage ratio with texture different colours image that is retrieved。

$STDC=\frac{SimilarNum}{M}---\left(19\right)$

Wherein, SimilarNum represent in the front M width image returning result be retrieved the image picture number with texture different colours, M has the total number of images order of identical texture different colours with the image that is retrieved in whole image library。The STDC measure algorithm accuracy rate to retrieving similar images, STDC value is more big, represents more forward with the image of the image similarity that is retrieved, and retrieval effectiveness is more good。Otherwise, retrieval effectiveness is more poor。

(3) identical image arrangement density D D (DistributionDensity)

DD represents with the image that is retrieved with texture with the image of color at the distributing position returned in image, before namely how many images comprise all with the image that is retrieved with the texture image with color。DD reflects the accuracy rate of retrieval, and DD value is more little, represents more forward with texture image distribution with color, and retrieval effectiveness is more good, and on the contrary, retrieval effectiveness is more poor。

The present invention tests and extracts LBP that (P, R) under equivalent formulations is (8,2) and retrieve as the characteristic vector of image, the image that image library 1 is spliced for texture primitive。For verifying the effectiveness of the inventive method, the method that the present invention is proposed with extract texture primitive by LBP conversion and carry out the method retrieved and the image search method based on texture primitive spatial distribution carries out contrast experiment。In table 1, the method that method one proposes for the present invention；Method two: image carries out LBP conversion, and statistics each row and each column LBP value are the frequency distribution of 0, and the frequency distribution cycle is the size of texture primitive, and the LBP characteristic vector extracting texture primitive is retrieved；Method three: based on the image search method of texture primitive spatial distribution。Adopt pertinency factor STSC (N=3), STDC and DD as the evaluation criterion of retrieval。

Table 1: retrieval result statistical table

STSC and DD is the evaluation criterion weighing identical image pertinency factor。From result statistical table 1 it can be seen that when retrieving identical image, the retrieval effectiveness of method one is best, and pertinency factor reaches 98.25%, and method two is taken second place, and pertinency factor is 92.98%, and method three effect is worst, and pertinency factor is 91.22%。The arrangement density value of method one is minimum, is 3.11, and method two is taken second place, and arrangement density value is 4.16, and the arrangement density value of method three is maximum, is 4.53。

Method one and method two are all that the content according to image self dynamically obtains texture primitive, texture primitive is extracted characteristic vector and retrieves。This search method avoids the diversity of the aspects such as the texture variations, amplitude of variation or the interval that produce because arrangement mode is different, is effectively increased the retrieval effectiveness with repetition texture primitive image, improves the pertinency factor of identical image。Method two extracts the accuracy rate of texture primitive lower than method one, thus pertinency factor is not as method one when retrieving identical image。Method three is by the size immobilization of texture primitive, but different types of image texture primitive is not of uniform size fixed identical, and texture primitive size immobilization can not accurately reflect the repeatability published picture as own content, and therefore pertinency factor is not so good as method one and method two。

When retrieving similar image, the retrieval effectiveness of method one is best, and accuracy reaches 61.88%, and method three is taken second place, and accuracy is 54.13%, and the retrieval effectiveness of method two is worst, and accuracy is 54.13%。The texture primitive accuracy of method two is lower than method one, and therefore, when retrieving similar image, accuracy is lower than method one。There will be the extraction incomplete situation of texture primitive when extracting texture primitive, now texture primitive can not correctly reflect the texture information that image is most basic, thus causing that retrieval rate reduces, causes that the accuracy of method two is lower than method three。

For verifying the stability of the inventive method, the method present invention proposed in different magnitude image libraries respectively contrasts with the image search method based on texture primitive spatial distribution, adopts pertinency factor (STSC) as the evaluation criterion of retrieval。Wherein, image library 2=is different from the image of the texture primitive+texture primitive splicing of the image zooming-out of image library 1；Ten thousand arbitrary images of image library 3=image library 1+1；Ten thousand arbitrary images of image library 4=image library 1+5。

Table 2: different magnitude image library retrieval result statistical tables

From table 2 it can be seen that when in different magnitude of image library, retrieval has the image of identical texture primitive, the method for the present invention is all much better than control methods。The inventive method can retrieve in image library the image with identical texture primitive。The first texture primitive according to picture material dynamic extraction image because of the inventive method, extract feature to texture primitive and retrieve, it is to avoid because of arrangement mode or comprise the picture material diversity that number difference causes。Method three can only retrieve texure primitive identical image when retrieval, but it is identical but comprise the image that number is different to retrieve texture primitive, and therefore pertinency factor is far below the inventive method。When retrieving identical texture primitive image, different magnitude of image library pertinency factor is essentially identical, it was demonstrated that the image that retrieval is had identical texture primitive by the method that the present invention proposes has certain stability。

Design patent retrieval method based on dynamic texture primitive provided by the invention, based on design patent, for different images texture primitive fixed equal feature not of uniform size, it is proposed to a kind of according to picture material dynamic extraction texture primitive and the image retrieval algorithm that texture primitive is extracted feature。When retrieval has the image of repetition texture primitive, the retrieval effectiveness of the present invention is better than the existing search method based on entire image, can retrieve and there is identical texture primitive, but the image that arrangement architecture is different, can be used for image texture structural similarity to judge, there is good retrieval effectiveness, it is possible to meet the needs of practical application well。

Embodiment described above only have expressed embodiments of the present invention, and it describes comparatively concrete and detailed, but therefore can not be interpreted as the restriction to the scope of the claims of the present invention。It should be pointed out that, for the person of ordinary skill of the art, without departing from the inventive concept of the premise, it is also possible to making some deformation and improvement, these broadly fall into protection scope of the present invention。Therefore, the protection domain of patent of the present invention should be as the criterion with claims。

Claims (3)

1. the design patent retrieval method based on dynamic texture primitive, it is characterised in that comprise the following steps:

Step 1): extract texture primitive；

Step 2): texture primitive is extracted the LBP feature that Corpus--based Method is analyzed, and the characteristic vector extracted is normalized and similarity measurement, return retrieval result。

2. the design patent retrieval method based on dynamic texture primitive according to claim 1, it is characterised in that described step 1) particularly as follows:

Step A: the image of texture primitive to be extracted is carried out the pretreatment operation such as gray processing, denoising, normalization；

Step B: image being done LBP conversion, extracts initial texture primitive, such as formula (3), formula (4), formula (5), shown in formula (6)；

Wherein: LBP converts as shown in formula (1) and formula (2):

If original image matrix I=(p_ij)_h×w, p_ijFor at (i, j) pixel value at place。Image array after LBP converts is I_LBP=(p '_ij)_h×w, p '_ijFor p_ijValue corresponding after doing LBP conversion。LBP conversion process such as formula (1):

$p_{ij}^{\&prime;}=\underset{l=0}{\overset{2^N-1}{\&Sigma;}}{2}^l\&times;S(g_l-g_c)---\left(1\right)$

$S\left(g_l-g_c\right)=\left\{\begin{array}{c}1,g_l\&GreaterEqual;g_c\\ 0,g_l<g_c\end{array}\right.---\left(2\right)$

Wherein, g_cCentered by some pixel value, i.e. p_ij, g_lFor neighborhood territory pixel value, N is the radius of neighbourhood。

$I_i=\left\{\begin{array}{c}True,abs\left(L_i-L\min \right)\&le;Lratio\&times;\left(L\max -L\min \right)\\ False,abs\left(L_i-L\min \right)>Lratio\&times;\left(L\max -L\min \right)\end{array}\right.---\left(3\right)$

H=min (I_i-I_i+1)(4)

$R_i=\left\{\begin{array}{c}True,abs\left(R_i-R\min \right)\&le;Rratio\&times;\left(R\max -R\min \right)\\ False,abs\left(R_i-R\min \right)>Rratio\&times;\left(R\max -R\min \right)\end{array}\right.---\left(5\right)$

W=min (R_i-R_j)(6)

Wherein, L_i, projection value that i ∈ [0, h-1] is horizontal direction, computing formula such as formula (7) and shown in formula (8)；R_j, projection value that j ∈ [0, w-1] is vertical direction, computing formula such as formula (9) and shown in formula (10)；

$L_i=\underset{j=0}{\overset{w-1}{\&Sigma;}}{L}_j^{\&prime;},0\&le;i\&le;h-1---\left(7\right)$

$L_j^{\&prime;}=\left\{\begin{array}{c}1,p_{ij}^{\&prime;}=Num\\ 0,p_{ij}^{\&prime;}\&NotEqual;Num\end{array}\right.---\left(8\right)$

$R_j=\underset{i=0}{\overset{h-1}{\&Sigma;}}{R}_i^{\&prime;},0\&le;j\&le;w-1---\left(9\right)$

$R_i^{\&prime;}=\left\{\begin{array}{c}1,p_{ij}^{\&prime;}=Num\\ 0,p_{ij}^{\&prime;}\&NotEqual;Num\end{array}\right.---\left(10\right)$

Step C: draw respectively to the right and downwards take onesize with initial texture primitive and not superimposed images block as sliding window；

Step D: judge initial texture primitive and draw whether take image block equal, as shown in formula (11)；If it is equal with initial texture primitive to only exist a sliding window, go to step E；If two sliding windows are all unequal with initial texture primitive, expand initial texture primitive, go to step C；If two sliding windows are all equal with initial texture primitive, go to step E；

$isBE(block1,block2)=\left\{\begin{array}{cc}1,& per\&GreaterEqual;threshold2\\ 0,& per<threshold2\end{array}\right.---\left(11\right)$

Wherein,

$per=\frac{\underset{i=0}{\overset{h-1}{\&Sigma;}}\underset{j=0}{\overset{w-1}{\&Sigma;}}isEqual\&lsqb;p_1\left(i,j\right),p_2\left(i,j\right)\&rsqb;}{H\&times;W}---\left(12\right)$

$isEqual\&lsqb;p_{i,j},p_{i,j}^{\&prime;}\&rsqb;=\left\{\begin{array}{c}1,abs\&lsqb;p_{i,j}-p_{i,j}^{\&prime;}\&rsqb;\&le;threshold\\ 0,abs\&lsqb;p_{i,j}-p_{i,j}^{\&prime;}\&rsqb;>threshold\end{array}\right.---\left(13\right)$

Wherein, p_{I, j}, p '_{I, j}Pixel value for different images block corresponding pixel points。

Step E: if the sliding window that there is certain direction is unequal with initial texture primitive, obtains new sliding window at direction slip certain distance in allowed band, go to step D；

Step F: using current initial texture primitive as image texture primitive。

3. the design patent retrieval method based on dynamic texture primitive according to claim 1, it is characterised in that described step 2) in extract LBP feature particularly as follows:

Use " equivalent formulations " that the schema category of LBP operator is carried out dimensionality reduction, as shown in formula (14):

${\mathrm{LBP}}_{P,R}^{riu2}=\left\{\begin{array}{cc}\underset{i=0}{\overset{P-1}{\&Sigma;}}S\left(g_i-g_c\right)& U\left({\mathrm{LBP}}_{P,R}\right)\&le;2\\ P+1& otherwise\end{array}\right.---\left(14\right)$

Wherein,

$U\left({\mathrm{LBP}}_{P,R}\right)=\left|S\right(g_{P-1}-g_c)-S(g_0-g_c\left)\right|+\underset{i=0}{\overset{P-1}{\&Sigma;}}|S\left(g_i-g_c\right)-S\left(g_{i-1}-g_c\right)|$ ,

If x >=0, S (x) takes 1, and otherwise S (x) takes 0, g_cRepresent the gray value of center pixel, g in neighborhood_iRepresenting the gray value of each pixel in the border circular areas that radius is R, wherein the span of i is 0 to P-1。