CN102385757A - Semantic restriction texture synthesis method based on geometric space - Google Patents

Abstract

The invention discloses a semantic restriction texture synthesis method based on geometric space, and the synthesis result has the structural characteristics different from the master drawing. The method has the key point that a texture topological structure is analyzed and processed, and a two-dimension texture grid is used for describing the particular characteristics of the texture. Firstly, pixel mask of the sample texture is used for forming a sample texture gird. Then the target texture grid is synthesized according to the restriction conditions of the given topological structure, namely the small blocks of grids given by a user. Finally, under the guide and control of the texture grid, each pixel of the target texture is synthesized by adopting a per-pixel synthesis algorithm based on characteristic vector, so that the aim of changing the texture topological structure can be realized. The invention realizes the aim of synthesizing a plurality of textures with different structural characteristics based on one sample texture.

Description

A kind of semantic constraint texture synthesis method based on geometric space

Technical field

The present invention relates to a kind of texture synthesis method based on master drawing.

Background technology

Synthetic based on the texture of master drawing is the very popular and important research direction in computer graphical and iconology field.In early days, people mainly study and how to utilize a fritter sample texture to obtain the bulk texture, have emerged synthetic and based on numerous outstanding methods such as the block-by-block of piece splicing thought are synthetic by pixel based on the neighborhood matching algorithm.Middle and later periods, particularly in recent years, a large amount of research is all no longer simple was devoted to promote synthetic quality and raising speed; But demonstrate diversity from many aspects; Synthetic like characteristic matching, dynamic texture synthetic, non-homogeneous texture is synthetic; Multi-dimension texture is synthetic, and the mixing of texture and level and smooth conversion etc.

1) planar grains synthetic method

By the at every turn synthetic pixel of pixel texture synthesis method, synthesize based on the neighborhood matching algorithm.This algorithm is shown in the imparametrization sampling synthetic method of Efros etc. the earliest, and basic thought is: the distance of the neighborhood through contrasting all master drawing neighborhood of pixels and pixel to be synthesized is chosen the optimum matching pixel, is so undertaken by pixel, thereby synthesizes target texture.Neighborhood of pixels is generally pressed vertical direction and horizontal direction structure, is divided into full neighborhood, half neighborhood, L shaped neighborhood.Full neighborhood is used for walking abreast, and texture synthesizes and the optimization of texture, and half neighborhood and L shaped neighborhood are suitable for the serial synthetic method.But parallel controllable synthesis method is a kind of exchange method in real time.Utilize the outward appearance vector to promote synthetic quality greatly based on the appearance space texture of this method is synthetic.Parallel controlledly synthesis is used for the mixing of structured image recently again by improvement such as Risser.The block-by-block synthetic method is chosen a texture block at every turn and is copied in the target texture from master drawing.Utilization dynamic programmings such as Efros etc. and Kwatra are sought a separatrix in the overlapping region of new piece and composite part.Wang Tiles texture synthetic technology is the basis with the method for Efros etc.; Have real-time, its newest research results is

etc. half randomness Tile synthetic method.

2) curved surface texture synthesis method

Synthetic texture mainly contains three kinds of modes at present on curved surface: first kind is directly synthetic, earlier with the curved surface refinement, calculates the color on all curved surface summits again.This mode is the most suitable dynamic texture synthetic.Second kind is indirect synthesis mode, and the method based on texture map of Ying etc. and Lefebvre etc. is divided into several zones with whole curved surface to be mapped on the plane, synthetic in the plane again these zone-textures.The third is to spread texture block to curved surface, till whole curved surface all is capped.In the method for Soler etc., the corresponding curved surface triangle of texture block.The method of Praun etc. repeats bedding with an irregular grain on curved surface.Fu etc. expand to Wang Tiles method on the curved surface based on the PolyCube mapping of Tarini etc.

3) signature analysis and matching process

Texture particle and feature distortion method have realized controlling texture is synthetic at the pel layer.Wu etc. utilize characteristic pattern to keep the architectural feature of texture.Based on effective characteristic analysis method, the mixing of texture just is achieved with level and smooth conversion.Because the existence of physical factors such as perspective and illumination, the texture in the photo often has very strong stereoscopic sensation.Liu etc. utilize deformation domain in photo, to carry out the texture replacement, and Eisenacher etc. have then realized from the synthetic texture of photo based on Jacobian field.Characteristic matching is synthetic based on the texture optimized Algorithm, the characteristic curve on synthetic Matching Model as a result surface.Dischler etc. carry out semantic relevant control based on texture grid to texture is synthetic.

Summary of the invention

Technology of the present invention is dealt with problems and is: texture can be divided into structural and the unstructuredness texture from architectural feature, can be divided into all even non-homogeneous texture according to the distribution situation of characteristic again.To homogeneous texture property texture, the present invention proposes a kind of semantic constraint synthetic method, through its semantic feature (architectural feature) is analyzed and handled, makes the user can select the architectural feature of synthetic texture by the wish of oneself.Having solved existing texture synthesis method can only synthesize with the sample texture and have the defective with a kind of texture of architectural feature.

Technical scheme of the present invention is:

(1) texture grid is synthetic

Based on the synthetic target texture grid of piece splicing.Generate a left margin and right margin coupling with sample texture grid (extraction or user are given from sample texture), the gridblock of coboundary and lower boundary coupling.A plurality of copies of this gridblock are spliced each other obtain the initial target texture grid, and then, reduce the multiplicity of target texture grid, promote mesh quality through optimizing.Here; Be divided into two kinds of situation: the one, be exactly the texture grid of sample texture if be used for the sample grid of synthetic target texture grid; The grid of target texture grid and sample texture has with a kind of topological structure so, and like this, the final objective texture just is consistent on semantic structure with master drawing; The 2nd, if synthesize with another sample grid that is different from the grid of sample texture; Then target gridding has different topological structures with the grid of sample texture; Thereby the final objective texture will have the semantic structure of user's appointment, no longer be subject to sample texture.

(2) under the constraint of texture grid, synthesize target texture based on the neighborhood matching algorithm

The present invention adopts traditional neighborhood matching algorithm can't obtain gratifying synthetic quality because will realize that the semantic constraint texture is synthetic, assists synthetic so introduce eigenvector.Prior synthesizing method is not intended to change the texture semantic feature, can directly construct neighborhood of pixels by vertical and horizontal direction, and in the present invention, the structure of neighborhood need be according to eigenvector.Eigenvector is by the decision of the relative position of pixel and textural characteristics.See eigenvector and primitive boundary cardinal principle vertical (Fig. 6) intuitively.

For target texture, earlier to the grid cell of each target texture grid, picked at random is one from all grid cells of the texture grid of sample texture, and the mapping relations of these two grid cell internal point are set up in mapping according to polygon.Each pixel that is target texture is in this way confirmed a position in sample texture.Here, this position coordinates can not just be an integer, can confirm pixel color based on bilinear interpolation.After the initialization pixel, according to the full neighborhood of eigenvector structure pixel, correct each pixel color based on the neighborhood matching algorithm again, obtain net result.

The present invention mainly contains 2 contributions: the first, provide a kind of 2 d texture grid composition algorithm, be applicable to irregular, approximate rule, and regular grid.Than the recursive optimization method of Dischler etc., our method does not need a large amount of manual inputs.The second, the use characteristic vector promotes synthetic effect, has solved traditional neighborhood matching algorithm and has been used for the unfavorable problem of semantic constraint texture synthetic effect, also is superior to existing sub-texture method.

Description of drawings

Fig. 1 algorithm overall flow figure;

Fig. 2 (a) sample texture, Fig. 2 (b) pel mask, Fig. 2 (c) characteristic pattern, Fig. 2 (d) texture grid;

Fig. 3 gridblock boundary matching cost;

The antidote of Fig. 4 gridblock;

Fig. 5 (a) is a sample grid, and Fig. 5 (b) is by the synthetic result of sample grid;

Fig. 6 (a) has marked the eigenvector of pixel in the sample texture, and Fig. 6 (b) has marked the eigenvector of pixel in the target texture;

Fig. 7 (a) is a sample texture, and Fig. 7 (b) is the existing methods synthetic effect, and Fig. 7 (c) is the synthetic effect of the inventive method;

Fig. 8 (a) is a sample texture, and Fig. 8 (b) is semantic constraint condition (each master drawing has two different constraint condition respectively), and Fig. 8 (c) is with same master drawing, the effect of synthetic texture under various boundary conditions.

Embodiment

Our method relies on texture grid, and it is the structure in a kind of two-dimensional geometry space, is used to describe the architectural feature of texture, is made up of summit and limit.This grid can obtain (Fig. 2) according to image processing techniques: pel mask (a kind of binary map that generates sample texture earlier; White portion is the pel district; Other zones are black); According to pel mask generating feature figure, each lines is that single pixel is wide in the characteristic pattern then, generates the texture grid of sample texture at last with characteristic pattern.This grid is taken as sample grid and is used for synthetic target texture grid.Because sample grid also can be artificially given, so which type of architectural feature the user just can have according to the wish decision target texture of oneself.

Semantic constraint texture synthesis method based on geometric space of the present invention is specific as follows:

First step: synthetic target texture grid

Our whole texture grid building-up process is the basis with gridblock boundary matching cost function.As shown in Figure 3, the computing method of boundary matching cost are: for any two gridblock p and q, b _pBe the right margin of p, b _qIt is the left margin of q.b _pTo intersect with the limit in some grids, changing a kind of saying promptly has some (can be 0) limits by border b _pCut apart, exist some limits equally by b _qCut apart.If by b _p, b _qThe number on the limit of cutting apart separately is different, then claims b _p, b _qDo not match, or the coupling cost be infinity, otherwise, claim that they mate each other, also claim restrained boundary each other.Suppose that at present they mate each other, order

Cut-point and b are pressed in expression _pThe i bar limit apart from the ascending order arrangement of initial end points (vertically upper extreme point is got on the border, and horizontal boundary is got left end point),

Be respectively two end points on limit.The in here is an endpoint attribute, representes this end points and b _pAffiliated gridblock is positioned at b _pHomonymy, otherwise out representes to be positioned at b _pHeteropleural, as shown in Figure 3.

Coordinate figure get with respect to b _pThe coordinate offset amount of initial end points.Equally,

Represent respectively by border b _qThe i bar limit of cutting apart, and two end points on this limit.Limit

and

are actually two line segments in the two-dimensional space, and we calculate the distance between them from both direction respectively.Calculated from the following formula

to

Distance:

$D(e_p^i,e_q^i)=(1+\mathrm{Dist}(v_p^{i,\mathrm{in}},v_q^{i,\mathrm{out}}))(1+\mathrm{\λ}||\stackrel{\→}{n}(v_p^{i,\mathrm{in}},v_p^{i,\mathrm{out}})-\stackrel{\→}{n}(v_p^{i,\mathrm{in}},v_q^{i,\mathrm{in}})|\left|\right)---\left(1\right)$

Wherein, is the distance of

; representes direction vector, and λ is a weight.

calculation coincided with opposite.We define the edge

and

The matching cost is:

$E(e_p^i,e_q^i)=D(e_p^i,e_q^i)+D(e_q^i,e_p^i)---\left(2\right)$

Existing hypothesis is respectively by border b _pAnd b _qThe limit number of cutting apart separately is N _Cut, the coupling cost on these two borders is so:

$E(b_p,b_q)=\underset{i=1}{\overset{N_{\mathrm{cut}}}{\mathrm{\Σ}}}E(e_p^i,e_q^i)---\left(3\right)$

For any gridblock MP; The set that four edges circle of remembering it is made up of them for

is B (MP), simultaneously this four edges circle restrained boundary and set thereof are designated as then always matees cost and be:

$E(B\left(\mathrm{MP}\right),\hat{B}\left(\mathrm{MP}\right))=\underset{u=t,b,l,r}{\mathrm{\Σ}}E(b_{\mathrm{MP}}^u,\widehat{b_{\mathrm{MP}}^u})---\left(4\right)$

Wherein It is the border

And border

The coupling cost, calculate with formula (3),

Corresponding b _p,

Corresponding b _qFollowing formula is called gridblock boundary matching cost function.

Based on the boundary matching cost function, we will synthesize and be divided into two stages: initialization procedure and optimizing process.

1.1) our target gridding (is designated as M _Tgt) initialization procedure is: (user is given or from sample texture, extract, and is designated as M by sample grid earlier _Smp) generating a seamless gridblock, a plurality of copies that again will this seamless gridblock are spliced into the initial target grid.The seamless gridblock here is meant that only need its copy be carried out simple combination just can obtain a complete grid, and does not need interpolation, deletion, mobile grid summit or limit.In the present invention, this special gridblock is designated as MP _s

Our strategy is: from all from M _SmpGridblock in the search " the best " gridblock (be designated as MP _o), its vertex position is corrected obtained MP again _sSo how to search for " the best " gridblock, this boundary matching cost function of the present invention just part that plays a role.For gridblock MP, we are with its coboundary

As lower boundary

Restrained boundary, with the restrained boundary of lower boundary, and with the given left margin of the same manner as the coboundary And right margin

Restrained boundary, go out MP with computes then _sThe border.

$B\left({\mathrm{MP}}_o\right)=\underset{\mathrm{MP}\∈\mathrm{\Ω}\left(M_{\mathrm{smp}}\right)}{\arg \min }\left(E(B\left(\mathrm{MP}\right),\hat{B}\left(\mathrm{MP}\right))\right)---\left(5\right)$

Ω (M wherein _Smp) be all M _SmpIn the set that constitutes of gridblock, arg min representes to make expression formula to reach the parameter value of minimum value.In theory, M _SmpThe gridblock that comprises is infinitely many.We limit any two gridblocks at M in experiment _SmpIn alternate position spike be the integral multiple of a constant, boundary length is also done this constraint, like this Ω (M _Smp) be exactly a finite set.Need to prove in addition: M _SmpWith the texture grid of sample texture be incoordinate, only if do not plan to change the semantic feature of target texture as the user, just at this moment both are equal to.

In most cases, MP _oBe not a seamless gridblock, need through correct it vertex position obtain MP _sThe vertex set of note arbitrary mess piece MP is V (MP), wherein is positioned at outside summit, MP border and constitutes set V _Out(MP), with V _Out(MP) any summit is adjacent and be positioned at inner summit, border and constitute set V in _In(MP).For V _Out(MP) and V _In(MP) any summit in, its new coordinate equal old coordinate add corresponding vertex coordinate in the restrained boundary and 1/2 (Fig. 4).Other summits constitute set V _Neu(MP), correct according to following formula:

$d\left(v^{\mathrm{neu}}\right)=\underset{k=1}{\overset{N_{\mathrm{in}}}{\mathrm{\Σ}}}{\mathrm{\ω}}_{k}d\left(v_k^{\mathrm{in}}\right)/\underset{k=1}{\overset{N_{\mathrm{in}}}{\mathrm{\Σ}}}{\mathrm{\ω}}_k,v^{\mathrm{neu}}\∈V_{\mathrm{neu}}\left(\mathrm{MP}\right)---\left(6\right)$

Wherein, N _InBe V _In(MP) number of vertex in,

Be V _In(MP) k summit in, v ^NeuBe V _Neu(MP) any summit in.

D (v ^Neu) be respectively

, v ^NeuThe coordinate offset amount.ω _kThe expression weights,

So d (v ^Neu) be exactly V _In(MP) weighted mean value of all apex coordinate side-play amounts in.Now, we just can use MP _sInitialization target texture grid.

1.2) repeat to splice obviously with a seamless gridblock and can cause too high multiplicity, impact effect is so what next will do is exactly to reduce target texture grid M _TgtMultiplicity, in other words be grid optimization.Our method is an iterative process, at every turn from M _TgtIn selected gridblock, again from M _SmpIn look for boundary length new piece replacement consistent with it.When treating replace block, in fact confirmed four constraint borders when selected, just can utilize formula (5) to obtain the best gridblock of replacing then and (be designated as MP _r).At last, with aforementioned summit antidote to MP _rBe spliced in the target gridding after correcting.

At M _TgtIn to select gridblock to be replaced, method of the present invention be that target gridding is divided into the unified rectangular block of size, to every from M _SmpK candidate's replace block of middle search also therefrom selected a replace block that conduct is final at random.Grid synthetic method effect of the present invention is as shown in Figure 5.

Second step: calculated characteristics vector and synthetic target texture under the grid constraint

Neighborhood of pixels in the neighborhood matching algorithm is a high dimension vector, has comprised r, g, the b component of all pixels in the neighborhood.If the size of neighborhood of pixels is 5*5, its dimension is exactly 75 so.In the prior synthesizing method, the neighborhood of each pixel all is to construct according to vertical direction and horizontal direction.Yet synthetic for the semantic constraint texture, because the architectural feature of target texture is different with the architectural feature of sample texture, can not simply adopt above-mentioned neighborhood make.For example, for the fragment of brick sample texture of rule, lines in its characteristic pattern or be vertical, or be level.Do this restriction now: the user hopes that the primitive shapes in the target texture to be synthesized is hexagonal.Certainly exist oblique line in the characteristic pattern of target texture at this moment.When adopting the neighborhood matching algorithm to synthesize target texture,, be prone to know that its neighborhood can not find the neighborhood of pixels of special coupling in master drawing for the pixel near the hexagon hypotenuse if still adopt traditional neighborhood make.In order to address this problem, we ask a direction to sample texture and target texture for each pixel, are called eigenvector.Near the pixel of hypotenuse, its direction is vertical substantially with hypotenuse, just can in master drawing, find the neighborhood of pixels of coupling according to the neighborhood of this directional structure vectorical structure; Before synthetic, all calculate its eigenvector for each pixel; With the fragment of brick sample texture is example, and easy knowledge pixel characteristic vector wherein all is vertically or level basically.And have the texture of regular hexagonal pel, then more than two kinds of the eigenvector of its pixel, but also be several discrete values.

We utilize characteristic distance generating feature vector according to following method in experiment.At first, in the image of white background, draw out texture grid with black.This image is carried out filtering with Gaussian filter (the Gaussian filter radius in the experiment is 5.0).At this moment, the r of the pixel color in this image (or g, b, because r=g=b) component value (span 0.0～1.0) is exactly the characteristic distance of pixel in the corresponding texture.The eigenvector of each pixel is its weighted mean value to the direction of all neighbor pixels (being designated as p), and weights are calculated as follows:

w(p)＝1/((f _p+e)d _p) (7)

Wherein, f _pIt is the characteristic distance of p.d _pBe the distance of p to the centre of neighbourhood, e is one and is used to avoid removing 0 decimal, as 0.001.When the structure neighborhood of pixels, the vertical direction in the corresponding classic method of eigenvector.

Building-up process (also claiming the rasterizing process) under the grid constraint is: calculate the characteristic distance of all pixels in sample texture and the target texture earlier according to texture grid, and the pixel in the target texture is pressed the characteristic distance ascending sort.According to this order each pixel basis eigenvector is constructed its neighborhood then.From master drawing, search for the optimum matching pixel again.Compare method better effects if of the present invention (Fig. 7) with the sub-texture grid method of Dischler etc.

The inventive method is applicable to regular veins, approximate rule texture, and irregular grain.The sample texture that we select for use in experiment has been included these three types.The method effect is as shown in Figure 8.Experimental situation is the PC of Inter Core (TM) i5 2.8GHz CPU, 4G internal memory, operation Windows 7 operating systems.Sample texture resolution is 128 * 128, and the figure elemental size surpasses 20 * 20.The texture grid generated time of target texture is approximately 15 seconds, and to different sample grid, the generated time amplitude of variation is within 5 seconds.Semantic constraint based on grid is synthetic, about 2 minutes consuming time.

Claims (3)

1. semantic constraint texture synthesis method based on geometric space comprises following two steps:

(1) utilize image processing techniques from sample texture, to extract or, then based on the synthetic target texture grid of gridblock splicing by the given sample texture grid of user;

(2) generate the eigenvector of sample texture and target texture according to sample texture grid and target texture grid, and adopt based on the synthetic target texture of the neighborhood matching algorithm of eigenvector.

2. a kind of semantic constraint texture synthesis method according to claim 1 based on geometric space, it is characterized in that: step (1) is specific as follows:

Synthesizing in geometric space of 2 d texture grid carried out, and the target texture grid is spliced mutually by several onesize gridblocks; For any two unidimensional gridblock p, q,, claim that so the right margin of p and the left margin of q are complementary if the number on the limit in the grid of being cut apart by the right margin of p equals the limit number cut apart by the left margin of q; The coupling cost function is then being input by two groups of limits of these two boundary segmentation; Every limit and its corresponding limit in second group to first group; Calculate their distance; Then with all these limits between obtain an arithmetic number apart from addition, be called the coupling cost of left margin of right margin and the q of p; For p, possibly there are a plurality of gridblocks, their left margin all is complementary with the right margin of p, and it is best to mate that minimum gridblock of cost so; Then through two steps; Can obtain the target texture grid: at first generate the gridblock that mate each other on one group of border according to gridblock boundary matching cost function and the given fritter grid of user; Said fritter grid also is called the semantic constraint condition, and this grid also can extract from texture; Concrete generative process is from the fritter grid, to select the plurality of grids piece, and mate each other on the border of these gridblocks; To each matching scheme, go out in this matching scheme the coupling cost between all coupling borders with the boundary matching cost function calculation, and add up and obtain total coupling cost of this matching scheme; Matching scheme the best that total coupling cost is minimum; The mutual splicing of the gridblock of optimum matching scheme has just obtained the target texture grid then.

3. a kind of semantic constraint texture synthesis method according to claim 1 based on geometric space, it is characterized in that: step (2) is specific as follows:

Employing is based on each pixel of the synthetic target texture of neighborhood matching algorithm of eigenvector, and the neighborhood of a pixel is to be a square area at center with this pixel; Neighborhood is equivalent to a high dimension vector, and building method is that the top left corner pixel from neighborhood begins to collect r, g, b component value by the sweep trace order by pixel; The distance of the neighborhood of neighborhood matching algorithm through contrasting all master drawing neighborhood of pixels and pixel to be synthesized is chosen the optimum matching pixel, and with this pixel color as color of pixel to be synthesized; So undertaken by pixel, thus synthetic target texture, said neighborhood matching algorithm based on eigenvector is when the structure neighborhood, and according to the eigenvector of pixel, the eigenvector of so-called pixel is a direction, and the hithermost primitive boundary of this direction and pixel is perpendicular.

Images