CN102637304A - Method for synthesizing isotropic/anisotropic texture on geometric surface based on GPU (Graphics Processing Unit) - Google Patents

Abstract

The invention discloses a method for synthesizing an isotropic/anisotropic texture on a geometric surface based on a GPU (Graphics Processing Unit), belonging to the field of computer graphics. The method comprises the following steps of: (1) selecting a texture image sample E and carrying out Gabor filtration on the image E to obtain a feature space of the E; (2) generating a grid peak structure S of the E; establishing a multiresolution pyramid structure according to the S to obtain a hierarchical grid model sequence, and carrying out isotropic surface texture synthesis; distributing a texture coordinate of a peak i according to a texture coordinate of a father peak in the grid model sequence on each layer of the pyramid structure; disturbing a texture coordinate of a geometric peak; carrying out iterative correction on the texture coordinate according to the feature space; and acquiring M nearest adjacent peaks of the grid peak; and (3) carrying out equidistant shift conversion on the texture coordinate of the peak and carrying out anisotropic surface texture synthesis. The method is high in synthesis efficiency and capable of keeping shape feature and deformation consistency.

Description

Geometric jacquard patterning unit surface isotropy/different in nature texture synthesis method based on GPU

Technical field

The present invention relates to a kind of geometric jacquard patterning unit surface isotropy/different in nature texture synthesis method that quickens through GPU; More particularly; The present invention relates to the programmable functions of a kind of GPU of utilization and the method for efficient processing performance three-dimensional geometry superficial makings aggregate velocity thereof, belong to field of Computer Graphics.The synthetic result of texture can be used for fields such as Film Animation, virtual reality, electronic game.

Background technology

1.1 the texture based on sample image is synthetic

In recent years, the synthetic easy property of map image details that significantly increased of texture to surface mesh.Algorithm early at random small images is sticked on three-dimensional surface; Use the alpha hybrid technology to hide interblock slit [E.Praun then; A.Finkelstein, and H.Hoppe.Lapped textures.Proceedings of SIGGRAPH 2000, pp.465-470].Recently, tailor's method [A.Efros and W.Freeman.Image quilting for texture synthesis and transfer.Proceedings of SIGGRAPH 2001, pp.341-346; L.Liang, C.Liu, Y.Xu, B.Guo, and H.Shum.Real-time texture synthesis by patch-based sampling.ACM Transactions on Graphics, vol.20, no.3, pp.127-150,2001; V. Kwatra, A.Schodl, I.Essa; G. Turk, and A.Bobick.Graphcut textures:Image and video synthesis using graph cuts.ACM Transactions on Graphics, vol.22; No.3, pp.277-286,2003; K.Zhou, P.Du, L.Wang; Y.Matsushita, J.Shi, B.Guo; And H.Shum.Decorating surfaces with bidirectional texture functions.IEEE Transactions on Visualization and Computer Graphics, vol.11, no.5; Pp.519-528,2005.] can produce better synthetic result, it reduces the uncontinuity of little block boundary through careful placement fritter texture.Place after these texture block [L. Liang, C.Liu, Y.Xu; B.Guo, and H.Shum.Real-timetexture synthesis by patch-based sampling.ACM Transactions on Graphics, vol.20; No.3, pp.127-150,2001; S.Magda and D.Kriegman; Fast texture synthesis on arbitrary meshes.ACM SIGGRAPH 2003 Sketches & Applications; 2003; P.1.] simply use the alopha hybrid technology to hide the texture slit, and [A.Efros and W.Freeman.Image quilting for texture synthesis and transfer.Proceedings of SIGGRAPH 2001, pp.341-346; K.Zhou, P.Du, L.Wang; Y.Matsushita, J.Shi, B.Guo; And H.Shum.Decorating surfaces with bidirectional texture functions.IEEE Transactions on Visualizationand Computer Graphics, vol.11, no.5; Pp.519-528,2005.] further strengthen the flatness of transition between the slit through searching for minimum shear gap.Because human vision is to the edge in the texture; Angle point is very responsive with other high-level characteristic, [Q.Wu and Y.Yu.Feature matching and deformation for texture synthesis.ACM Transactions on Graphics, vol.23; No.3; Pp.364-367,2004.] from sample texture, extract characteristic pattern, the method for use characteristic coupling and image deformation is come optimum maintenance characteristic continuity then.Other associated texture synthetic method comprise based on pixel and based on the mixed method of piece strategy [A.Nealen and M.Alexa.Hybrid texture synthesis.Proceedings of the 14th Eurographics workshop on rendering.Eurographics Association, 2003, p.105.]; Object in the picture is carried out texture synthetic [H.Fang and J.Hart, Textureshop:texture synthesis as a photograph editing tool.ACM Transaction on Graphics, vol.23 again; No.3, pp.354-359,2004.]; Parallel controlled texture synthetic [S.Lefebvre and H.Hoppe.Parallel controllable texture synthesis.ACM Transactions on Graphics, vol.24, no.3 that GPU is last; Pp.777-786,2005.], [S.Lefebvre and H. Hoppe.Appearance-space texture synthesis.ACM SIGGRAPH 2006; P.548.] and use texture synthesis method [the V. Kwatra of expectation minimization optimized Algorithm; I.Essa, A.Bobick, and N.Kwatra.Texture optimization for example-based synthesis.ACM SIGGR APH2005; P.802.] .Huang et al. [H.Huang; X.Tong, and W.Wang.Accelerated parallel texture optimization.Journal of Computer Science and Technology, vol.22; No.5; Pp.761-769,2007.] on GPU, realized K-similar search and principal component analysis (PCA) (PCA), and further quickened the texture optimization of back.The Gabor wave filter has desirable optimal properties for texture analysis; Their direction and radial frequency bandwidth are adjustable, allow to unite resolution optimization in space and the spatial frequency.Bovid and Clark [A.Bovik; M.Clark; And W.Geisler.Multichannel texture analysis using localized spatial filters.IEEE Transactions on Pattern Analysis and Machine Intelligence; Pp.55-73,1990.] proposed to use 2D Gabor wave filter to carry out the computing method that visual texture is analyzed.They find that this method is all very suitable for cutting apart of artificial and natural texture, as in expecting.Nearest Gilet and Dischler [G.Gilet; J.Dischler, et al.An Image-Based Approach for Stochastic Volumetric and Procedural Details.Computer Graphics Forum, vol.29; No.4; Wiley Online Library, 2010, pp.1411-1419.] use the Gabor wave filter to calculate the texture similarity so that carry out playing up based on image.

1.2 superficial makings is synthetic

The texture synthesis method of three-dimensional surface in the ten years development in the past, has made the complicated image detailed design of arbitrary surface become significantly and has been more prone to.First kind method [G.Turk.Texture synthesis on surfaces.Proceedings of the28th annual conference on Computer graphics and interactive techniques.ACM New York; NY; USA, 2001, pp.347-354.; L.Wei and M.Levoy.Texture synthesis over arbitrary manifold surfaces.Proceedings of the 28th annual conference on Computer graphics and interactive techniques.ACM New York; NY; USA; 2001, pp.355-360.; X.Tong, J.Zhang, L.Liu, X.Wang; B.Guo, and H.Shum.Synthesis of bidirectional texture functions on arbitrary surfaces.ACM Transactions on Graphics, vol.21; No.3, pp.665-672,2002.; S.Zelinka and M.Garland.Interactive texture synthesis on surfaces using jump maps.Proceedings of the 14 ^ThEurographics workshop on Rendering.Eurographics Association Aire-la-Ville, Switzerland, Switzerland; 2003; Pp.90-96.] realized based on each sampling of pixel imparametrization [A.Efros and T.Leung.Texture synthesis by non-parametric sampling.International Conference on Computer Vision, vol.2, no.9; 1999, pp.1033-1038.; L.Wei; And M.Levoy.Fast texture synthesis using tree-structured vector quantization.Proceedings of the27th annual conference on Computer graphics and interactive techniques.ACM Press/Addison-Wesley Publishing Co.New York; NY, USA, 2000; Pp.479-488.] texture is synthetic, and needs extra mapping or resampling so that play up.Comparatively speaking, some method is used the original texture image in the process of playing up, and uses texture mapping hardware to carry out the index of this image.A kind of method early wherein is people such as Soler [C.Soler; M.Cani, and A.Angelidis.Hierarchical pattern mapping.Proceedings of the 29th annual conference on Computer graphics and interactive techniques.ACM New York, NY; USA; 2002, pp.673-680.], they realize the establishment of fritter through the cluster of tri patch.Zelnika and Garland [S.Zelinka and M.Garland.Jump map-based interactive texture synthesis.ACM Transactions on Graphics; Vol.23; No.4; Pp.930-962,2004.] also expanded their jump mapping, so that use in the coordinate of original image index on each summit.Lefebvre and Hoppe [S.Lefebvre and H.Hoppe.Appearance-space texture synthesis.ACM SIGGR APH 2006; P.548.] use hardware to realize being used for the extra memory storage of texture look-up table, and their method has utilized GPU to provide mutual texture synthetic speed.High-quality among their the synthetic result has partly cause to be because used line unit's mask and the PCA of neighborhood of pixels is analyzed.People such as Han [J.Han, K.Zhou, L.Wei; M.Gong, H.Bao, X.Zhang; And B.Guo.Fast example-based surface texture synthesis via discrete optimization.The Visual Computer, vol.22, no.9; Pp.918-925,2006.] having proposed the use expectation-maximization algorithm comes the gauging surface texture, and they have realized this algorithm on GPU.Method [E.Praun, A.Finkelstein, and H.Hoppe.Lapped textures.Proceedings of SIGGRAPH 2000, pp.465-470 before the great majority; G.Turk.Texture synthesis on surfaces.Proceedings of the 28th annual conference on Computer graphics and interactive techniques.ACM New York; NY; USA, 2001, pp.347-354.; L.Wei and M.Levoy.Texture synthesis over arbitrary manifold surfaces.Proceedings of the 28th annual conference on Computer graphics and interactive techniques.ACM New York; NY; USA; 2001, pp.355-360.; X.Tong, J.Zhang, L.Liu, X.Wang; B.Guo, and H.Shum.Synthesis of bidirectional texture functions on arbitrary surfaces.ACM Transactions on Graphics, vol.21; No.3, pp.665-672,2002.; J.Zhang; K.Zhou; L.Velho; B.Guo, and H.Shum.Synthesis of progressively-variant textures on arbitrary surfaces.ACM SIGGRAPH 2003, pp.295-302.] all extract the local neighborhood grid through local pressing and surface resampling technology.

1.3 texture

Be used in the actor model of the simplification that be everlasting game engine or 3D play up with the texture technology that the problem that a curved surface is flattened into a plane is closely related.Since Catmull [E.Catmull.A subdivision algorithm for computerdisplay of curved surfaces.1974.] proposes texture; A large amount of methods [E.Bier and K.Sloan.Two-part texture mappings.IEEE Computer Graphics and applications has appearred in document; Vol.6; No.9, pp.40-53,1986; P.Heckbert.Survey of texture mapping.IEEE Computer Graphics and Application, vol.6, no.11, pp.56-67,1986; S.Haker, S.Angenent, A.Tannenbaum; R.Kikinis, G. Sapiro, and M.Halle.Conformal surface parameterization for texture mapping.IEEE Transactions on Visualization and Computer Graphics; Vol.6; No.2, pp.181-189,2000; L. Wang, X. Gu, K.Mueller; And S.Yau.Uniform texture synthesis and texture mapping using global parameterization.The Visual Computer; Vol.21, no.8, pp.801-810; 2005.]. because most of 3D surface can not be launched, in last composograph, there is the texture distortion.And although the 3D surface can be launched, many methods also can cause distortion.Based on surface parameterization so that the texture synthesis method that flattens a 3D surface need resolve into discrete plane picture piece with a surface; And use the method that segmentation is shone upon to set up the corresponding relation in the isomorphic equivalent thing in 3D grid and their the 2D plane, and to minimize the distortion of this introducing through linear and nonlinear solution.At document [A.Sheffer; E.Praun; And K.Rose.Mesh parameterization methods and their applications.Foundations and Trends in Computer Graphics and Vision, vol.2, no.2; P.171, provided the good summary in the surface parameter method recently 2006.].

1.4 the problem that exists in the said method

Superficial makings has increased the authenticity of geometric model and rich.Although arbitrary surfaces is carried out texture synthetic problem by extensive studies, real-time and concurrency remain the main bottleneck of interactive applicationization.

The superficial makings synthetic method of introducing above can mainly be divided into two types: texture is synthetic with the texture that directly carries out from the teeth outwards.The former, for example UV mapping and UVW mapping unavoidably have significant texture distortion and heavy manually-operated shortcoming.Therefore, there have been a lot of research work to synthesize about directly carrying out texture at three-dimensional surface.Superficial makings is continuous and directly defines from the teeth outwards, although their forms with discrete single summit in reality show.Current superficial makings synthetic method is directed to static surface and often uses global parameterized, and this will cause texture slit or distortion.As everyone knows, realize that the mapping in surface parameter territory has very high complicacy, most typical, the surface parameterization that produces a low distortion needs considerable user intervention.

In addition, if texture will reflect the shape information on surface, consistent with significant geometry or deformation so is very important.Although the effects that controllability, well-designed vector field can let the synthetic generation of the texture of three-dimensional surface much change.But the not too many problem of paying close attention to the texture gradual change in body form or the geometric deformation situation of present these texture synthesis methods that propose.A lot of real object surfaces textures all have dimensional variation character.So in the time of our synthetic such article surface vein, we should produce the texture of this yardstick gradual change to produce more real texture synthetic effect.For example the zebra model all is the striped of large scale at health and the back of zebra usually, yet is very little at the yardstick of this striped of leg area.In order in virtual reality, to produce this real synthetic effect, this texture gradual change must be carried out careful consideration and design in the synthetic process of texture.Be not only in static model, for the model of deformation, we also want in the process of deformation, to keep the consistance of texture very much.Existing method does not also have well to solve the demand of foregoing description, can not produce real texture roll-off characteristic.

Parallel texture composition algorithm [L. Wei and M.Levoy.Order-independent texture synthesis.Computer Science Department, Stanford University, 2002 at classics; S.Lefebvre and H.Hoppe.Parallel controllable texture synthesis.ACM Transactions on Graphics (TOG), vol.24, no.3, pp.777-786,2005.] in, they do not consider the variation of yardstick field.

Summary of the invention

The objective of the invention is to solve the three-D grain surface and directly carry out isotropy, the synthetic problem of different in nature texture.It is synthetic that the inventive method utilizes programmable functions and the efficient processing power thereof of GPU to quicken texture.

Hypothetical trellis surface S=(V, C) with m * m texture sample image E (m * m is the size of unity and coherence in writing sample image, i.e. number of pixels), wherein V representes that grid vertex gathers, C representes the communication information between the summit.The synthetic mapping function S:V → E (each grid vertex that is about among the two-dimensional grid S is mapped to the coordinate among the two dimensional image E) that can think to calculate each summit of superficial makings; For each summit i ∈ V; Wherein S [i] representes the coordinate u among the E, so the color of summit i is E (u)=E [S [i]].Receive parallel texture synthesis strategy [S.Lefebvre and H.Hoppe.Appearance-space texture synthesis.ACM SIGGRAPH 2006 Papers.ACM, 2006, p.548; L. Wei and M.Levoy.Order-independent texture synthesis.Computer Science Department, Stanford University, 2002; S.Lefebvre and H.Hoppe.Parallel controllable texture synthesis.ACM Transactions on Graphics (TOG); Vol.24; No.3, pp.777-786,2005.] inspiration of thought; We are applied to arbitrary surfaces with the synthesis flow (promptly parallel texture synthesis strategy) of up-sampling-shake-correction, to reach the purpose that keeps geometric properties.Simultaneously, we have adopted the Gabor characteristic to improve the accuracy of K-Coherence search.

Technical scheme of the present invention is:

A kind of geometric jacquard patterning unit surface isotropy based on GPU/different in nature texture synthesis method the steps include:

1) chooses a texture image sample E and image E carried out Gabor filtering, obtain the Gabor feature space of image E;

2) the grid vertex structure S of generation image E; S sets up the multiresolution pyramid structure according to the grid vertex structure, obtains the grid model sequence of stratification, and the superficial makings that carries out isotropic is synthetic; Promptly each layer of pyramid structure is carried out:

A) superficial makings up-sampling distributes the texture coordinate of the summit i texture coordinate according to its father vertex in the grid model sequence;

B) superficial makings coordinate disturbance is carried out disturbance to the texture coordinate on how much later summits of up-sampling;

C) according to the Gabor feature space to the texture coordinate iteration correction;

D) for each grid vertex, obtain its M nearest adjacent vertex, be designated as V _p(p=1...M);

3) carry out the not skew conversion of equidistance for the vertex texture coordinate, the superficial makings that carries out anisotropic is synthetic.

Further, adopt the K-Coherence searching method to calculate the distance of said Gabor feature space; Wherein, the K-Coherence method has comprised K coordinate that has the pixel of similar neighborhood with current pixel for each sampled pixel makes up a coordinate set in the coordinate set.

Further, S resamples to surface mesh, and the summit of surface mesh is evenly distributed.

Further, adopt the canonical method of sampling that surface mesh S is resampled.

Further, (Y, Cb Cr) describe the color texture of image E, then to the brightness Y channel filtering of image E, thereby said Gabor feature space T are projected to the subspace (F of L dimension with brightness space _i) _I=1..L(R, G B) merge into H dimensional feature space (R, G, B, a F with the color space of said L n-dimensional subspace n and image E at last _i), i=1...L.

The method of further, carrying out the texture coordinate iteration correction according to the Gabor feature space is: the grid that at first the neighborhood triangle set of grid vertex is converted to (2N+1) * (2N+1) through projection; Each sampled point obtains a H dimension Gabor feature space by the interpolation relation then, makes the neighborhood N of each grid vertex i _iThe last Gabor proper vector of dimension of a H * (2N+1) * (2N+1) that forms; N is less than 5 integer greater than 1.

Further, said step 2) in, the superficial makings synthetic method of carrying out isotropic is: when synthesizing pixel p, calculate the candidate pixel collection of current point p according to the coupling of the pixel similarity in the Gabor feature space; Then,, concentrate also joining candidate pixel with k-1 the most similar pixel of its neighborhood in the sample texture for each pixel that candidate pixel is concentrated, at last from the concentrated pixel of mating the most found out of candidate pixel with the p neighborhood as final synthetic result.

Further, adopt section projection way, the neighborhood triangle set of grid vertex is converted to the grid of (2N+1) * (2N+1) through projection to grid vertex.

Further, adopt k-coherence texture matching process to carry out the neighborhood search on geometric jacquard patterning unit surface summit.

Further, the number of vertices of said pyramid network adjacent two layers is 4: 1.

Further; In the said step 3), at first obtain the not pair-wise offset yardstick field of its texture coordinate, squint for texture coordinate then according to the curvature on image pattern E surface; Obtain the superficial makings characteristic with the geometric jacquard patterning unit surface variation, the superficial makings that carries out anisotropic is synthetic.

Further, in the said step b), through two-dimentional Hash random function H:Z → [0,1] ²And the stray parameter by user's control comes the disturbance of control surface texture coordinate.

2.1Gabor space

It is synthetic based on the texture of sample image to be similar to 2D, and the similarity of the neighborhood of synthetic pixel and the neighborhood of sample image pixel is weighed through the calculated amount of characteristic distance usually.The pixel that finds the sample image similar with the synthetic surface pixel is to improve the key of the synthetic texture visual quality in surface.Method before (is also referred to as expressive space [S.Lefebvre and H.Hoppe.Appearance-space texture synthesis.ACM SIGGRAPH 2006Papers.ACM through spatial neighborhood characteristic or dimensionality reduction moral spatial neighborhood characteristic; 2006, p.548; ]) distance carry out the coupling of neighborhood.Yet the multi-scale characteristic of sampled pixel is that SSD (Sum-of-Squared-Difference, the difference of two squares distance) distance that is difficult to through the texture space neighborhood is caught.

The present invention obtains the Gabor feature space of E through pending image E is carried out Gabor filtering.Because the frequency and the multiscale space analysis of textural characteristics have been unified in Gabor filtering.

To 2D sample image E (x, y), its Gabor wavelet transformation is:

E _m，n(x，y)＝∑ _x1∑ _y1E(x ₁，y ₁)g _m，n(x-x ₁，y-y ₁). (1)

Wherein, g _{M, n}Expression document [B.S.Manjunath and W.-Y.Ma.Texture features for browsing and retrieval of image data.IEEE Trans.Pattern Anal.Mach.Intell.; Vol.18, no.8, pp.837-842; 1996.] in the self similarity bank of filters; That m, n represent respectively is yardstick and direction (m=1...C, the n=1...K) (x of Gabor wave filter in the document ₁, y ₁) expression geometric jacquard patterning unit surface arbitrary summit, the implication of N is the size of Gabor filter sample window.Filtering image amplitude average and variance are used for defining as follows respectively in the structural attitude space:

Filtering image amplitude average: $u_{m,n}=\frac{1}{N}{\mathrm{\Σ}}_x{\mathrm{\Σ}}_y\left|E_{m,n}(x,y)\right|,---\left(2\right)$

With

Filtering image amplitude variance: ${\mathrm{\σ}}_{m,n}=\sqrt{\frac{1}{N}{\mathrm{\Σ}}_x{\mathrm{\Σ}}_y{\left(\right|E_{m,n}(x,y)|-{\mathrm{\μ}}_{m,n})}^2},---\left(3\right)$

A lot of Texture classification technology; Adopt document [B.S.Manjunath and W.-Y.Ma.Texture features for browsing and retrieval o fimage data.IEEE Trans.Pattern Anal.Mach.Intell.; Vol.18; No.8, pp.837-842,1996.] use 4 yardsticks (C=4) and 6 directions (K=6).The Gabor feature space is following:

T＝[μ ₀， ₀σ _0，1μ _0，1…μ _3，5σ _3，5」(4)

Color texture can use the various colors space to describe, (R, G, B) or luminance/chrominance color space (Y, Cb, Cr).In texture composition algorithm of the present invention, the Gabor bank of filters is used for to the brightness Y channel filtering of sample texture image.Use PCA (Principle-Component-Analysis) method to project to the subspace (F of 6 dimensions the eigenvector of 48 dimensions _i) _I=1..6, also can project to L dimension (general dimension is less than 20), the space after the corresponding merging can be a H dimension (H=L+3).Therefore, the Gabor feature space that is used to synthesize can be expressed as 9 dimension spaces (R, G, B, F _i), i=1...6.The Gabor feature space of low dimension produces according to the method for original sample image through precomputation.

In order to quicken the texture aggregate velocity, used the K-Coherence searching method in the distance of Gabor feature space.The K-Coherence method has comprised K coordinate that has the pixel of similar neighborhood with current pixel for each sampled pixel makes up a coordinate set in this coordinate set.Experiment shows that the similar collection of sampled pixel can be found accurately, because the Gabor wave filter has the characteristic than better distinguishing characteristic of SSD index and texture structure.The primary structure and each the submodule call relation that have shown our methods at Fig. 1.Compare with other texture synthesis method, in the process that K-Coherence makes up, we have used the Gabor feature space, have kept texture and structural characteristic better more accurately, and are as shown in Figure 2.Fig. 3 has proved that equally also use Gabor characteristic can access the synthetic result of comparatively ideal texture.

2.2GPU pre-service

Carry out real-time superficial makings synthetic before; We utilize ReMesh [M.Attene and B.Falcidieno.ReMESH:An interactive environment to edit and repair triangle meshes.in IEEE International Conference on Shape Modeling and Applications; 2006.SMI 2006; 2006, pp.41-41.] method resamples to surface mesh S, and the summit of surface mesh is distributed as far as possible evenly.The grid of hierarchical structure is 4: 1 method realization equally through the number of vertices of controlling adjacent two layers.The set membership of the adjacent two layers of gold tower network can be formulated as P in the algorithm _L+1(j)=k, wherein j ∈ S _L+1, k ∈ S _l, and

$k=\arg {\min }_{k^{\′}\∈S_l}\left|\right|j-k^{\′}\left|\right|---\left(5\right)$

Wherein, what function p () represented is for the summit on the sub-grid, finds the solution the process on its corresponding father summit on father's grid; J is expression sub-grid S _L+1Any summit, k representes the summit (summit k is the father of summit j) of father's grid, S _lBe sub-geometric grid S _L+1Father's grid.Network S and sample image E concern one to one; S is a series of grid sequences from coarse to fine, E also be a series of texture sample (being pyramid structure) set membership from coarse to fine obtain through pre-service and preserve so as in real time synthetic acceleration up-sampling step.

We use Gauss's storehouse that Gaussian filter makes up the texture sample image.At the i stack layer, the radius of wave filter is 2 ^I-1+ 1.Through the processing that texture sample is done, the gaussian pyramid that obtains texture sample is expressed, and the texture sample of such gaussian pyramid structure is corresponding one by one with the pyramid structure of geometric grid for the back texture is synthetic.

2.3 the superficial makings of isotropic is synthetic

Because we regard superficial makings as vertex color and directly are synthesized to surface mesh, actual synthetic elementary cell is the curved surface summit.In work in the past, a texture pyramid from coarse to fine can be caught the texture structure and the little community of multiresolution.We are based on the mesh of vertices structure (network on summit is S) of curved surface; Set up a multiresolution pyramid structure (list of references: L. Wei and M.Levoy.Texture synthesis over arbitrary manifold surfaces.Proceedings of the 28th annual conference on Computer graphics and interactivetechniques.ACM New York by thick and thin 2D texture sample; NY; USA; 2001, pp.355-360.), to obtain the multi-resolution framework information in the 2D texture sample; We have also set up the multiresolution pyramid structure based on the mesh of vertices S of geometric jacquard patterning unit surface simultaneously; Set up the grid model sequence of stratification, and the multiresolution pyramid structure of combined with texture sample carries out by last from down, multiresolution from coarse to fine is synthetic.We use, and the stratification texture is synthetic to be created from low to higher resolution.In initial coarse resolution grid, we are through specify texture coordinate at random, and this is equivalent to the input albefaction noise texture at random at the pyramid top.For each pyramidal level, we carry out 3 composition algorithm steps: up-sampling, disturbance and correction.Each grid vertex is carried out and the irrelevant synthesis step of order, specifies as follows:

(1) superficial makings up-sampling

In the up-sampling of superficial makings, the texture coordinate of summit i is distributed by direct texture coordinate according to its father vertex (being in the thicker grid resolution) in the grid hierarchical model.The grid vertex algorithm here is unlike the up-sampling algorithm of 2 d texture.Because new algorithm is indifferent to the similarity of texture coordinate on direction and distance between the father and son summit here.Its reason is that the offset coordinates of texture coordinate between the father and son summit hierarchical relationship of computational geometry surface mesh is loaded down with trivial details, and directly inherits the visual quality of last comprehensive coordination present equivalence.So the Mesh on geometric grid surface promotes to sample and can is by mathematical expression in the new algorithm:

S _l＝(S _l-1[P _l(i)])modm

The texture coordinate that is summit i is distributed by direct texture coordinate according to its father vertex (being in the thicker grid resolution) in the grid hierarchical model.

(2) superficial makings coordinate disturbance

The factitious regularity of distribution appears in the effect synthetic for fear of texture, and the forcing function that we use coordinate carries out necessary disturbance for the texture coordinate on how much later summits of up-sampling.A forcing function J like this ₁(i) can pass through two-dimentional Hash random function H:Z → [0,1] ²(Z is exactly a two-dimensional coordinate number at random, H:Z → [0,1] ²What represent is to generate one 2 dimension coordinate at random, and two dimension scopes of coordinate are all in [0,1].) and a stray parameter by user control control.In the up-sampling of geometric jacquard patterning unit surface texture coordinate; Because the direct succession of coordinate between the father and son summit; The summit topological relation of its irregular geometric jacquard patterning unit surface has produced certain texture randomness; So under normal conditions, perturbation process is a kind of mutual treatment step, can be by the randomness disturbance of user's selection and the superficial makings of controlling.

(3) texture coordinate iteration correction

In the synthetic process of whole superficial makings, the core of superficial makings is the texture coordinate iteration correction process of parallelization.The texture trimming process of new algorithm comprises three steps: (1) each surface vertices i; We obtain 5 * 5 sampled point grids to each grid vertex on its section; (2) each sampled point can be obtained the Gabor proper vector of one nine dimension, the neighborhood N of each grid vertex i like this by the interpolation relation _iCan form the Gabor proper vector of one 9 * 5 * 5 dimension on (5 * 5 sampled point grids).(3) we obtain its M (M is 6) individual nearest adjacent vertex here for each grid vertex, are designated as V _p(p=1...6).(4) in reality was synthetic, we carried out k-coherence texture matching process the neighborhood search on geometric jacquard patterning unit surface summit.When synthesizing pixel p; Calculate the candidate pixel collection of current point p earlier according to the coupling of the pixel similarity in the Gabor feature space; Then; For each pixel that candidate pixel is concentrated, concentrate (implication of k is the neighbor pixel number of each pixel) also joining candidate pixel with k-1 the most similar pixel of its neighborhood in the sample texture, at last from the concentrated pixel of mating the most found out of candidate pixel with the p neighborhood as final synthetic result; The synthetic technology of this superficial makings has been considered the synthetic concurrency of texture and has been generated quality, can obtain Rapid Realization at GPU easily.

2.4 anisotropic superficial makings is synthetic

In new algorithm, we can synthesize the geometric jacquard patterning unit surface texture through adopting the not pair-wise offset on the different directions, thereby the texture that reaches anisotropic is apparent, shows the first ratio effect of different lines in the geometric jacquard patterning unit surface texture.Though it should be noted that the directly plane parameter effect of evaluation rule of traditional Jacobian field J; Its curve surface definition to the overall situation but is difficult to carry out accurate suitable parameters calculating; Particularly the vector field rotational transform of its local surfaces also has very big ambiguousness (E.Praun; A.Finkelstein, and H.Hoppe.Lapped textures [C] //Proceedings of SIGGRAPH ' 2000, Computer Graphics; Annual Conference Series, 2000:465-470).Though also can adopt geodesic distance to come the Jacobian field on definition set surface, the subjectivity that still has height is calculated in the rotational transform of its how much local surfaces.For this reason; New algorithm does not adopt the method for definition Jacobian field to synthesize the geometric jacquard patterning unit surface texture in anisotropic texture is synthetic; But seek geometric jacquard patterning unit surface is resolved to the anisotropic skew of neighbours' texture coordinate in the texture cell: for this reason, we are through the skew N for summit and its neighborhood texture coordinate _i(Δ) realized the texture synthetic effect with the Jacobi's transformation equivalence.We have defined not equidistance scaling field specifically $D_s=\left(\begin{array}{c}{D}_u\\ D_v\end{array}\right),$ Its Jacobi's transformation offset coordinates can be represented to be expressed as like this:

$N_i^{\′}\left(\mathrm{\Δ}\right)={\mathrm{Jacobian}}_i^{-1}{N}_i\left(\mathrm{\Δ}\right)=\left(\begin{array}{c}{D}_u\left(i\right)N_i{\left(\mathrm{\Δ}\right)}_u\\ D_v\left(i\right)N_i{\left(\mathrm{\Δ}\right)}_v\end{array}\right)$

Wherein, the u direction off-set value of Du (i) expression summit i in texture coordinate space (uv space); The v direction off-set value of Dv (i) expression summit i in texture coordinate space (uv space); Ni representes the local geometric neighborhood at summit i place, is that small one by one plane constitutes; N _i(Δ) is illustrated on the local small neighbourhood of summit i, the sampling point position of the direct neighbor of summit i; Ni ' (Δ) then representes, after Jacobi's skew, and the sampling point position of the direct neighbor of the part of i on the summit; Especially, if D _u=D _v, this shows the pair-wise offset that is scaled of each summit i, such being provided with can let the user generate the texture appearance that metamorphosis takes place with the geometric jacquard patterning unit surface change in location.In the texture of equidistance did not synthesize, we also can be provided with D simultaneously _u≠ D _v, the variation of this surperficial scaling field can cause texture cell, i.e. the form generation crimp effect of line unit.

Compared with prior art, good effect of the present invention is:

But our method adopts the pattern of parallel computation fully; Through making up two pyramid structures of geometric model surface and texture sample, utilize up-sampling, disturbance; And the operation of proofreading and correct three levels; Realized the synthetic effect quick true to nature of geometric jacquard patterning unit surface texture, in synthetic process with model in the synthetic order on each summit have nothing to do, realize so be well suited for GPU.For the synthetic superficial makings that can show style characteristic and geometric deformation, the yardstick field that changes surfacewise is that very big effect is arranged.Algorithm of the present invention can produce the texture that keeps consistency with geometric jacquard patterning unit surface through making up the yardstick field of reflection surface geometry characteristic or man-machine interactively design, can utilize Jacobi's transformation can regulate the different scale effect that superficial makings generates simultaneously.The texture of having that effect has the texture features that meets natural forms, perhaps reaches the style characteristic of design.The algorithm that the present invention proposes has following advantage:

1) algorithm directly synthesizes texture on the 3D surface, and does not need heavy manually-operated and overall surperficial parametrization.And we have attempted using the Gabor characteristic to extract the K-Coherence coupling of image.

2) because algorithm can carry out the texture of isotropy/opposite sex to be synthesized, so we can keep style characteristic and deformation consistance in synthetic texture.

3) because of algorithm flow and summit sequence independence, thus its parallelization fully, and on GPU, realize.

Description of drawings

Synthetic primary structure module of the parallel superficial makings of Fig. 1 and mutual relationship synoptic diagram

Fig. 2. the neighborhood resampling grid synoptic diagram of parallel processing grid vertex i;

Fig. 3. isotropic texture synthetic effect, test model have the individual summit in the individual summit, 400,000 (left sides) and 100,000 (right sides), memory space 431.5MB and 109.9MB video memory. and neighborhood resampling resolution is 7 * 7;

Fig. 4. anisotropic surface texture synthetic effect;

Fig. 4 (a). based on the heterogeneous surface scaling field synoptic diagram of surface curvature;

Fig. 4 (b). generate anisotropic superficial makings.On the camel model surface of test, the line unit of its body part is greater than the line unit size of its foot and neck;

Fig. 5. the present invention can be used in the synthetic effect synoptic diagram of superficial makings of deformable bodies.

Embodiment

Below in conjunction with embodiment the inventive method is described further.

The microcomputer of present embodiment is configured to Intel Core Duo 2.6G CPU, 1G internal memory, GeForce8800 video card and Window XP operating system.Generate the texture building-up process of geometric jacquard patterning unit surface according to following step:

Pretreatment stage:

(1), pending image E is generated based on the feature space of the Gabor wave filter coupling benchmark as sample texture according to as shown in Figure 1.

Be similar to based on the synthetic idea of calculating of the 2 d texture of sample, an important foundation stone of the texture synthetic technology among the present invention is exactly the measurement to pixel similarity in the space, and this measurement often is the basis with the tolerance of the characteristic distance between pixel.The suitable sample of the superficial makings that will in image E, search and synthesize finds the most appropriate texture block coupling according to similarity, is vital for improving visual quality.In the algorithm in the past, for the neighbourhood who is complementary through measuring distance in color space (be so-called appearance space reference: S.Lefebvre and H.Hoppe.Appearance-space texture synthesis.ACM SIGGRAPH2006, p.548).These technology often are difficult to obtain the texture feature of multiresolution, particularly are difficult to catch its space, distinctive texture neighbourhood for irregular or semicircular canal texture sample then.

(2) generate the neighborhood grid on surface mesh summit among the corresponding surface mesh S of E and to carry out texture synthetic.

In the building-up process of texture; At first adopt section projection way to grid vertex; Convert the set of the neighborhood triangle of grid vertex to 5 * 5 or 7 * 7 neighborhoods (generally be the grid at (2N+1) * (2N+1), N is generally greater than 1 less than 5 integer) resampling grid (as shown in Figure 2) through projection.Thereby carry out the parallelization texture processing for the neighborhood grid on each summit further thus; Generally speaking; Set up three-layer network compartment from coarse to fine time model earlier for geometric jacquard patterning unit surface; Then through on every layer of grid model, adopting up-sampling, the texture coordinate on the process generation model surface of disturbance and iteration correction.It is synthetic that such synthetic technology can be carried out isotropic texture quickly and easily, and can realize parallel the acceleration through graphic hardware, and Fig. 3 has provided the synthetic synoptic diagram as a result of relevant isotropic superficial makings.

(3) anisotropic texture composition algorithm and textured surface texture generate.

In the synthetic process of texture,, just can obtain anisotropic texture synthetic effect through iteration optimization if carry out the not skew conversion of equidistance for the vertex texture coordinate.As shown in Figure 4, at first obtain the not pair-wise offset yardstick field of its texture coordinate according to the curvature of camel model surface, squint for texture coordinate then, will obtain the superficial makings characteristic that changes with geometric jacquard patterning unit surface.Same this anisotropic texture composition algorithm can be applied to animation and deformation geometry surface, obtains realistic texture synthetic effect (as shown in Figure 5).

The GPU-CUDA boost phase:

New algorithm can fully operate in the graphic hardware platform CUDA framework of parallelization at the synthetic algorithm of superficial makings.We form the 1D/2D array to the structure of grid data with the indicator index matrix, manage as global memory's data of CUDA, by indicator index.We recomputate the synthetic data of data in the time of also need being stored in the GPU internal memory operation, comprise the grid relevant two groups of data with sample of being correlated with, and are described below respectively:

1. geometric grid data-grid vertex, surface vertices neighborhood vertex set, summit partial top point coordinate system, (normal, tangent vector), k nearest-neighbors vertex set, Jacobi's parameter on each summit.This local coordinate system is by vertex scheme vector, and primary tangent direction and subtangent direction be as z, x, and the y axle, what is called is meant that normally vectorial the and tangent vector of vertex scheme all carry out unitization.

2. sample data-Gauss's storehouse texture collection, k-coherence vertex set, the array of indexes of synthetic texture coordinate.

In the data organization in the CUDA program, we are for any summit, surface, generate geometrical vicinity (neighborhood) at its subrange, then we to carry out fast texture to this neighborhood piece synthetic.From the experiment of algorithm, with the surface on independently integration engineering and the grid.Because run time kernel is synthetic; The process that the texture of multipass is proofreaied and correct is in 5 * 5 or 7 * 7 summit neighborhood; It is synthetic to walk abreast, and promptly to any summit among the surface mesh S, in 5 * 5 or 7 * 7 summit neighborhood, proofreaies and correct; It is synthetic to adopt the method for step 1-5 to walk abreast then, and execution in step is described below:

1) the proper vector summit, neighbourhood on the current summit of collection; Proper vector summit, neighbourhood is confirmed by the nearest some summits of Euclidean distance.

2) apex coordinate that 2 of calculating rings are adjacent is on the local coordinate plane; Promptly calculate the feature space distance of each candidate point on the current summit of feature space middle distance.

3) all candidate samples point sets on collection and treatment summit;

4) calculate the feature space distance of each candidate point in feature space;

5) calculate the minor increment collection of above-mentioned distance, and obtain new map index.

In principle, each parallel processing process should possess identical CUDA operation kernel in these synthesis steps, but this restriction has often reduced the dirigibility that the GPU internal memory uses.According to actual needs, we are divided into a plurality of CUDA with whole process and handle kernel, the part of functions that each kernel treatment surface texture is synthetic.Specifically, each kernel is by 1 dimension Gabor feature space data, and perhaps one dimension K-coherence vector data is stored ephemeral data simultaneously in overall video card internal memory.Then, we call new kernel function repeatedly proper vector and k-coherence algorithm are got texture coordinate to the end.Make us can handle bigger geometric grid, big summit neighborhood scope at the parallel texture synthesis program that the CUDA exploitation realizes.

According to said process, select for use three models to carry out the compound experiment of superficial makings in this instance.In the table 1 three mold surface texture generated times are added up, and GPU is realized comparing with the CPU implementation efficiency.

The time efficiency statistics of table 1 superficial makings composition algorithm (CPU, GPU)

Table 1 has been explained new superficial makings composition algorithm respectively in GPU and CPU environment, and to various geometric model texture synthetic effects, each model has three grid levels under the resolution.Can know by table 1, adopt the parallel processing of GPU to quicken multiple greatly between 20-30 times.

In the present invention, we have proposed the synthetic technology of a kind of parallelization same sex based on GPU/different in nature superficial makings, have created the texture synthetic effect consistent with geometric properties, have removed the global parameterized computation process of grid simultaneously from.New technology is introduced the external appearance characteristic that the Gabor feature space is caught texture, thereby realizes the better maintenance for texture and structural characteristic.We have proposed the texture yardstick field of the same sex/opposite sex, can produce multiple different surface texture form effect, whole texture building-up process simple controllable.Texture shape and surface geometry have good consistance, and can be used for grain effect under the various geometry deformations.And the superficial makings synthetic technology of the opposite sex can be used for multiple static state and animation body surface.The synthetic method that the present invention utilizes GPU to quicken superficial makings can directly promote the use of in relevant animation and the geometry designs figures processing procedure, has stronger Practical significance.

Claims (12)

1. the geometric jacquard patterning unit surface isotropy based on GPU/different in nature texture synthesis method the steps include:

1) chooses a texture image sample E and image E carried out Gabor filtering, obtain the Gabor feature space of image E;

2) the grid vertex structure S of generation image E; S sets up the multiresolution pyramid structure according to the grid vertex structure, obtains the grid model sequence of stratification, and the superficial makings that carries out isotropic is synthetic; Promptly each layer of pyramid structure is carried out:

A) superficial makings up-sampling distributes the texture coordinate of the summit i texture coordinate according to its father vertex in the grid model sequence;

B) superficial makings coordinate disturbance is carried out disturbance to the texture coordinate on how much later summits of up-sampling;

C) according to the Gabor feature space to the texture coordinate iteration correction;

D) for each grid vertex, obtain its M nearest adjacent vertex, be designated as V _p(p=1...M);

3) carry out the not skew conversion of equidistance for the vertex texture coordinate, the superficial makings that carries out anisotropic is synthetic.

2. the method for claim 1 is characterized in that adopting the K-Coherence searching method to calculate the distance of said Gabor feature space; Wherein, the K-Coherence method has comprised K coordinate that has the pixel of similar neighborhood with current pixel for each sampled pixel makes up a coordinate set in the coordinate set.

3. according to claim 1 or claim 2 method is characterized in that surface mesh S is resampled, and the summit of surface mesh is evenly distributed.

4. method as claimed in claim 3 is characterized in that adopting the canonical method of sampling that surface mesh S is resampled.

5. the method for claim 1 is characterized in that (Y, Cb Cr) describe the color texture of image E, then to the brightness Y channel filtering of image E, thereby said Gabor feature space T are projected to the subspace (F of L dimension with brightness space _i) _I=1..L(R, G B) merge into H dimensional feature space (R, G, B, a F with the color space of said L n-dimensional subspace n and image E at last _i), i=1...L.

6. method as claimed in claim 5 is characterized in that the method for carrying out the texture coordinate iteration correction according to the Gabor feature space is: the grid that at first the neighborhood triangle set of grid vertex is converted to (2N+1) * (2N+1) through projection; Each sampled point obtains a H dimension Gabor feature space by the interpolation relation then, makes the neighborhood N of each grid vertex i _iThe last Gabor proper vector of dimension of a H * (2N+1) * (2N+1) that forms; N is less than 5 integer greater than 1.

7. like claim 1 or 5 or 6 described methods; It is characterized in that said step 2) in; The superficial makings synthetic method of carrying out isotropic is: when synthesizing pixel p, calculate the candidate pixel collection of current point p according to the coupling of the pixel similarity in the Gabor feature space; Then,, concentrate also joining candidate pixel with k-1 the most similar pixel of its neighborhood in the sample texture for each pixel that candidate pixel is concentrated, at last from the concentrated pixel of mating the most found out of candidate pixel with the p neighborhood as final synthetic result.

8. method as claimed in claim 6 is characterized in that adopting the section projection way to grid vertex, the neighborhood triangle set of grid vertex is converted to the grid of (2N+1) * (2N+1) through projection.

9. like claim 1 or 5 or 6 described methods, it is characterized in that adopting k-coherence texture matching process to carry out the neighborhood search on geometric jacquard patterning unit surface summit.

10. the method for claim 1, the number of vertices that it is characterized in that said pyramid network adjacent two layers is 4: 1.

11. the method for claim 1; It is characterized in that in the said step 3); At first obtain the not pair-wise offset yardstick field of its texture coordinate according to the curvature on image pattern E surface; Squint for texture coordinate then, obtain the superficial makings characteristic with the geometric jacquard patterning unit surface variation, the superficial makings that carries out anisotropic is synthetic.

12. the method for claim 1 is characterized in that in the said step b), through two-dimentional Hash random function H:Z → [0,1] ²And the stray parameter by user's control comes the disturbance of control surface texture coordinate.

Images