09.01.2015 Views

Estimation of surface normals for use in 3D rendering ... - IEEE Xplore

Estimation of surface normals for use in 3D rendering ... - IEEE Xplore

Estimation of surface normals for use in 3D rendering ... - IEEE Xplore

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

METHODS<br />

This study compared the gray-level<br />

gradient, the eigenvector and the<br />

modified eigenvector methods <strong>for</strong><br />

determ<strong>in</strong><strong>in</strong>g the <strong>surface</strong> <strong>normals</strong> and hence<br />

the <strong>surface</strong> rendered images.<br />

For the gray-level gradient method two<br />

adjacent pixels <strong>for</strong> each axis were <strong>use</strong>d<br />

<strong>in</strong> the calculation; the normal be<strong>in</strong>g<br />

N = (dx, dy, dz)<br />

where dx, dy, dz are the central<br />

differences and N is normalized to a unit<br />

vector.<br />

For the eigenvector method either 2, 4,<br />

or 6 neighbors <strong>in</strong> each axis direction and<br />

the gray scale values <strong>of</strong> these pixels<br />

were <strong>use</strong>d. The <strong>surface</strong> normal is<br />

N = V1 = (XI, yl, 21) '<br />

where V1 is the normalized eigenvector<br />

associated with the largest eigenvalue,<br />

11, <strong>of</strong> the covariance matrix A.<br />

For the modified eigenvector method all<br />

<strong>of</strong> the pixel above a threshold <strong>in</strong> a cubic<br />

neighborhood <strong>of</strong> the pixel under<br />

consideration and the re 1 at i ve<br />

coord<strong>in</strong>ates <strong>of</strong> these pixels, rather than<br />

their gray scale values, are <strong>use</strong>d to f<strong>in</strong>d<br />

the normal. The normal is<br />

N = V3 = (xl, yl, zl)<br />

where V3 is the normalized eigenvector<br />

associated with the smallest eigenvalue,<br />

h, <strong>of</strong> the covariance matrix A.<br />

The above three methods to compute<br />

the <strong>surface</strong> normal were evaluated on data<br />

sets conta<strong>in</strong><strong>in</strong>g a spherical object. Two<br />

<strong>of</strong> the data sets had a step transition<br />

between the object and the background,<br />

one with noise and one without. The<br />

other two sets had a l<strong>in</strong>ear transition<br />

between the object and the background,<br />

one with noise and one without. The<br />

techniques also were tested on human CT<br />

data acquired us<strong>in</strong>g a Picker<br />

International, Inc. CT scanner.<br />

RESULTS<br />

Both the eigenvector method and the<br />

modified eigenvector method produced<br />

images with comparable quality to the<br />

gray-level gradient method. The<br />

quantitative studies showed that the<br />

eigenvector method was more accurate than<br />

the gray-level gradient method when 4 or<br />

6 neighbors were <strong>use</strong>d <strong>in</strong> the calculation<br />

and that the modified eigenvector method<br />

was more accurate than the gray-level<br />

gradient method and the eigenvector<br />

method <strong>for</strong> the step transition between<br />

the object and the background. The<br />

eigenvector method was more accurate than<br />

the modified eigenvector method when the<br />

transition between the object' and the<br />

background was l<strong>in</strong>ear. For both the<br />

eigenvector method and the modified<br />

eigenvector method calculations us<strong>in</strong>g 4<br />

neighbors were more accurate than us<strong>in</strong>g 2<br />

neighbors.<br />

The studies us<strong>in</strong>g human CT data<br />

showed that the eigenvector method<br />

revealed some f<strong>in</strong>e structures that could<br />

not be seen us<strong>in</strong>g the gray-level gradient<br />

method; however, this method tended to<br />

exaggerate the small structures. The<br />

modified eigenvector method produced<br />

images that looked smooth, but flat.<br />

Both the eigenvector method and the<br />

modified eigenvector method were<br />

computationally more expensive than the<br />

gray-level gradient method.<br />

REFERENCES<br />

[lIM. Magnusson, R. Lenz, P.E. Danielsson<br />

(1988). Evaluation <strong>of</strong> Methods <strong>for</strong> Shaded<br />

Surface Display <strong>of</strong> CT-Volumes.<br />

Proceed<strong>in</strong>gs <strong>of</strong> the 9th ICPR, 1287-1294.<br />

121 B.T. Phong (1975) . Illum<strong>in</strong>ation <strong>for</strong><br />

Computer Generated Pictures.<br />

Communications <strong>of</strong> the ACM, 18(6) :311-317.<br />

[SlU. Tiede, K.H. Hoehne, M. Bomas, A.<br />

Pommert, M. Riemer, G. Wiebeck (1990).<br />

Investigation <strong>of</strong> Medical <strong>3D</strong>-Render<strong>in</strong>g<br />

Algorithms. <strong>IEEE</strong> Computer Graphics and<br />

Applications, 10:41-53.<br />

[4]C.H. Wood, X. D<strong>in</strong>g, W. Satt<strong>in</strong> (1991).<br />

Eigenvector Surface Normal <strong>Estimation</strong> <strong>in</strong><br />

Medical Imag<strong>in</strong>g. 9th Annual Meet<strong>in</strong>g <strong>of</strong><br />

the Society <strong>for</strong> Magnetic Resonance<br />

Imag<strong>in</strong>g, Chicago, IL.

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!