Sei sulla pagina 1di 6

CWP-655

Visualization of 3D tensor elds derived from seismic images


Chris Engelsma & Dave Hale
Center for Wave Phenomena, Colorado School of Mines, Golden, CO 80401, USA

(a)

(b)

Figure 1. A 3D seismic image with a traced layer displaying the tensors as ellipsoids (a), and removing that slice shows the 3-dimensional structure of the tensors from that layer (b).

ABSTRACT

In image processing, tensors derived from seismic images are used as parameters in procedures such as structure-oriented smoothing. Visualizing these tensors allows us to qualitatively assess their computation and construction. We describe a computationally eective technique to render these tensors as ellipsoid glyphs. Key words: seismic visualization ellipsoid tensor

INTRODUCTION

Visualizing tensor elds has always been a complicated task. While there are many ways in which scientists can visualize scalar or vector elds, displaying tensor elds in an intuitive manner remains a challenge. In recent years, a number of techniques have been proposed discussing methods for displaying 3D tensors. In the medical industry, the continuity of tensor elds is emphasized by constructing hyperstreamlines or streamtubes for diusion MRI tensors (Delmarcelle and Hesselink, 1993; Jianu et al, 2009). In stress evaluation, the eects of a tensor eld on a given media have been visualized through bending mesh volumes, simulating the eects

of a stress tensor to demonstrate anisotropic deformation (Zheng and Pang, 2002). In geophysics, tensors are being used to help guide seismic horizon tracing (H ollt et al, 2009). Recently, a number of methods have been proposed for displaying 3D tensors. For the purposes of image processing in exploration geophysics, tensors are often derived directly from the images. While a method such as hyperstreamlining would give insight into the continuity of the tensor eld, it is also advantageous to render each tensor individually by depicting them in an intuitive manner as Figure 1 demonstrates. These discrete representations of tensors in 3D are called glyphs. Glyph representation can take many forms, and the benets of choosing dier-

188

C. Engelsma & D. Hale

(a)

(b)

(c)

Figure 2. A scalar eld (a), a vector eld (b), and a tensor eld (c). The visualization complexity increases dramatically with the number of quantities represented at each node.

ent shapes have been explored (Kindlmann, 2004). However, for the purpose of this paper, each tensor glyph is represented as an ellipsoid. Tensors help increase the ecacy of image processing by guiding the orientation of the operation. This is the principle behind structure-oriented smoothing (Hale, 2009). Because these tensors are used as parameters in dierent processing techniques, we must determine their accuracy. We therefore wish to explore a method of visualizing these parameters in a discretized manner, allowing us to evaluate any arbitrary tensor. We describe an algorithm to visualize tensors derived from seismic images, and demonstrate methods for evaluating the tensors authenticity. By displaying tensors as ellipsoid glyphs, this visualization method provides an intuitive and interactive method for relating the tensors directly back to the image. We also expedite the rendering process by making our method computationally fast and ecient.

(van Vliet and Verbeek, 1995; Fehmers and H ocker, 2003), which are smooth outer-products of image gradients. The eigen-decomposition of a 3D structure tensor S(x) yields: S = u uuT + v vvT + w wwT , (1)

where the eigenvalues of u , v , and w are sorted so that u v w 0. (2)

TENSOR GEOMETRY

The challenge with visualizing tensors stems from their multivariate nature. With scalar elds, each sample is a representative of one number. Vector elds follow the same concept, but each point is now represented by three numbers in 3D. Tensor elds introduce another step in intricacy because we are now representing six unique numbers at every point in space. A visual representation of this increasing complexity is shown in Figure 2. Simultaneous visualization of six numbers extends beyond conventional visualization techniques unless we understand the geometry of the tensors. 2.1 Metric tensor eld D(x)

By convention, u is dened as the eigenvector that traverses the direction of the largest gradient. In a 3D seismic image, this typically refers to the direction perpendicular to geologic layering. Both eigenvectors v and w tend to lie in the plane of locally planar features in the image. We then compute anisotropic metric tensors D(x), using a process outlined by Hale (2009), whereby we compute image semblances. We choose the eigenvectors of each metric tensor D to be the same as those for the corresponding structure tensor S. The dierence between D and S lie only in their eigenvalues. Specically, the eigen-decomposition on D(x) is D = s3 uuT + s2 vvT + s1 wwT , (3)

where we construct eigenvalues s1 , s2 , and s3 such that 0 s3 s2 s1 1. (4)

An ideal structure-oriented procedure which honors the dominant structural features of our image (e.g. rock bedding layers and faults), requires a tensor eld that accurately represents these features. This is accomplished by rst computing the structure tensors S(x)

Our metric tensor D(x) is a 3 3 symmetric, positivedenite matrix. The largest eigenvalue s1 , corresponding to the eigenvector w, is semblance computed within a locally linear (1D) set of voxels aligned with w. Each eigenvalue s2 , corresponding to the eigenvector v, is semblance computed within a locally planar (2D) set of voxels orthogonal to the corresponding eigenvector u. (The plane orthogonal to u contains the eigenvectors v and w). Finally, each eigenvalue s3 represents

Visualization of 3D tensor elds

189

u u

xT Ax = 1
Figure 3. A non-axis-aligned ellipsoid.

v v

w w

AV = VD
Figure 4. Eigenvalue and eigenvector relationship to the three principle radii of an ellipsoid. Each axis within the ellipsoid is equal to the eigenvector divided by the square root of their eigenvalues.

semblance computed for a locally spherical (3D) set of voxels. 2.2 Ellipsoid glyphs

We consider the denition of an ellipsoid to be xT Ax = 1, (5)

ACCELERATED RENDERING

where A is a square symmetric positive-denite matrix, and x is any point along the surface of the ellipsoid which satises this equation (note Figure 3). We can likewise dene a unit sphere in the same manner by replacing matrix A with the identity matrix I. Equation 2 provides a useful denition because it describes an ellipsoid that is not axis-aligned; the eigenvectors of A are arbitrarily aligned in space. Considering the denition of eigenvector orthonormality, we dene the eigenvectors as the three principle axis radii, and the inverse of the square root of the eigenvalues as their respective sizes (see Figure 4) (Strang, 2003). This geometric relationship enables us to construct the tensors as ellipsoids; in a computer, these are illustrated as glyphs. 2.3 Geologic analogy

Constructing each glyph requires computing the location of roughly one thousand vertices to be used in a triangle mesh (see Figure 5). This computation becomes costly when one begins to display a large set of ellipsoids throughout a 3D survey. We therefore expedite the rendering process. If we rst compute the vertex locations for a unit circle, we then obtain the desired ellipsoid by applying the appropriate matrix transformations. This greatly reduces the computational cost. We compare the equations of a unit circle, xT x = 1, to our desired transformed ellipsoid coordinates, yT Ay = 1. We also dene the eigen-decomposition of A to be A = VDVT , (8) (7) (6)

Ellipsoid glyph representations of metric tensors demonstrate the local orientation of the image. In particular, for a perfectly horizontal layer, we expect the ellipsoid to be oblate, because u > v w . Likewise, for complete isotropy within an image, we expect our ellipsoid to be a sphere (u = v = w ). The local geologic orientation of the formation is also reected, so ellipsoids incorporate the same strike and dip qualities as their corresponding locations in the seismic image. Because the intention of displaying these ellipsoids is to qualitatively assess the how accurately the tensors have been constructed, we expect that they follow the bedding layers in the image. This allows us to judge the veracity of image processing techniques guided by these tensors.

where V is a 33 orthogonal matrix containing the eigenvectors of A stored as column vectors, and D is a diagonal matrix storing eigenvalues u v w . We now replace A in equation 8 with equation 7, and we get yT VDVT y = 1. (9)

Given the property of a diagonal matrix that D = 1 1 D 2 D 2 , we expand equation 9 to get yT VD 2 D 2 VT y = 1.


1 1

(10)

Equation 10 is the equation of an ellipsoid in terms of

190

C. Engelsma & D. Hale

(a)

Figure 6. Ellipsoids selected to follow a single layer by point-and-click method. The strike and dip of the local formation is apparent.

ordinates for each ellipsoid. The vertex coordinates for a unit sphere are only computed once and stored.

IMPLEMENTATION METHODS

Here we discuss two methods for overlaying ellipsoid glyphs on 3D seismic data. Both methods oer dierent techniques to visualize tensors, and both may be used for dierent investigative purposes. We show two approaches to displaying tensor elds: a point-and-click and an axis-aligned panel method. 4.1
(b) Figure 5. Two glyphs: a unit sphere (a) and an ellipsoid (b). The ellipsoid was computed using matrix transformations on the spheres mesh.

Point-and-click method

the eigenvectors and eigenvalues of A. From equation 6, we observe that the variable x represents the coordinates of a unit sphere. To transform the sphere into an ellipsoid, we derive our desired coordinates y in terms of our computed coordinates x: y = VD 2 x.
1

(11)

V represents a rotation matrix which realigns the princi1 ple axes of the unit sphere. The matrix D 2 is a nonuniform scaling matrix containing the inverse of the square root of the eigenvalues. Performing equation 11 is more computationally ecient than explicitly computing each vertex because this process passes 12 numbers to the graphics card instead of recalculating one thousand co-

Figure 6 shows ellipsoids that are selected along a given layer. This is performed by a succession of mouse clicks which place an ellipsoids center on the sample nearest to the cursor. Note that every ellipsoid appears oblate with varying thicknesses, and that each ellipsoid has a dip that reects the local orientation. Focusing on the shape of the ellipsoids is important for determining whether or not the tensor eld has been correctly computed. Note also that the ellipsoids appear in 3D relative to the image slice. This allows the user to rotate freely, preserving the location and visibility of the tensor. In a similar way, the user can drag the cursor along the image and observe the changes in the ellipsoids at each point in space. By not sticking the ellipsoids as in Figure 6, the user can watch the tensor mold to the layers and identify discrepancies this way. 4.2 Axis-aligned panels

Placement of ellipsoids along an axis-aligned panel (see Figure 7) shows an overall distribution of tensor clusters. This process involves discretizing tensors along a

Visualization of 3D tensor elds

191

Figure 7. A panel of tensor ellipsoids. Each ellipsoid represents a metric tensor, and is equally sampled along the x-axis panel. The closeup emphasizes the variation in shape as well as angle of each ellipsoid at each point in space.

3D seismic panel, allowing the user to qualitatively assess many tensors simultaneously. From a macroscopic viewpoint, this will enable the user to grab a broad perspective of the underlying structure. While Figure 7 shows ellipsoids attached to a single panel, displaying ellipsoids on all three axis-aligned panels is a reasonable interpretation method as well. The caveat of this approach is that ellipsoids will not necessarily fall directly on a point of interest. Because the ellipsoids are evenly sampled along the panel, the user is only permitted to see tensors that lie on that sampling interval. For a more detailed survey of tensor ellipsoids, the point-and-click approach is more eective.

CONCLUSIONS

Displaying tensor elds is an ongoing topic of research in the eld of visualization. The inherent problem with displaying tensors is due to the amount of information contained in each sample. In geophysical applications of image processing, tensors are derived from the seismic images in order to design structure-oriented operations. For the purpose of quality assessment, we choose to display these tensors as glyphs shaped as ellipsoids. Constructing ellipsoids from tensors works in our favor, as our metric tensors t this geometric relationship. By performing an eigen-decomposition of the tensor matrix, we obtain three orthonormal eigenvectors

and their corresponding eigenvalues, which can be represented as the three principle axis directions and their corresponding radii. Expediting the process involves precomputing the vertices of a unit sphere, and performing both a rotation and scaling matrix. Because our tensors are derived from the seismic image, we show that the shape of the ellipsoid relates to the local orientation of the image. Flat layers yield discshaped, oblate ellipsoids, and isotropic environments are more spherical. Ellipsoids must have the same strike and dip of the surrounding area. We demonstrate two methods of displaying tensor elds: the rst dynamically selects ellipsoids at a clicked voxel; the second involves discretizing ellipsoids along an axis-aligned panel. Both techniques provide intuitive visualization of the geologic substructure, with pointand-click placement allowing for detailed investigation of a specic point.

ACKNOWLEDGEMENTS

We thank the Rocky Mountain Oileld Testing Center, a facility of the U.S. Department of Energy, for providing us with the data used for Figures 1, 6, and 7. We also thank our colleagues for their continual discussion and feedback on this topic. We also thank Diane for her very helpful editing of this paper.

192

C. Engelsma & D. Hale

REFERENCES
Delmarcelle, T. and L. Hesselink, 1993, Visualizing secondorder tensor elds with hyperstreamlines: IEEE Computer Graphics and Applications, 13, 2533. Fehmers, G.C. and C.F.W. H ocker, 2003, Fast structural interpretations with structure-oriented ltering: Geophysics, 63, 12861293. Hale, D., 2009. Structure-oriented smoothing and semblance: CWP Report 635, 261270. H ollt, T. et al, 2009, Seismic horizon tracing with diusion tensors: Presented at IEEE VisWeek 2009. Jianu, R., C . Demiralp, and D.H. Laidlaw, 2009, Exploring 3D DTI ber tracts with linked 2D representations: IEEE Transactions on Visualization and Computer Graphics, 15, 14491456. Kindlmann, G., 2004, Superquadric tensor glyphs: In Proceedings of the Joint Eurographics - IEEE TCVG/EG Symposium on Visualization 04, 147154. Strang, G., 2003, Introduction to linear algebra: WellesleyCambridge Press, 553. van Vliet, L.J. and P.W. Verbeek, 1995, Estimators for orientation and anisotropy in digitized images: Proceedings of the rst annual conference of the Advanced School for Computing and Imaging ASCI 95, 442450. Zheng, X. and A. Pang, 2002, Volume deformation for tensor visualization: Proceedings of the conference on Visualization 02, 379386.

Potrebbero piacerti anche