Neural Networks in Computer Graphics


Semi-Automated Segmentation

One of the most tedious jobs in medical image processing is hand-drawing the boundaries around tissue of interest. We explored ways of training neural networks to duplicate the decisions made by a human anatomist while the human is tracing boundaries, then letting the neural network complete the tracing, with corrections from the human when necessary. A prototype, in MATLAB, of a complete system for neural-net-assisted tracing of region contours and the assembly into 3-dimensional models is described in this final report from our 1997-98 CASI grant.

Stew Crawford completed his Ph.D. dissertation in this area in 2003:

  • Crawford-Hines, S. (2003) Machine Learned Boundary Definitions for an Expert’s Tracing Assistant in Image Processing, Ph.D. Dissertation, Department of Computer Science, Colorado State University, Fort Collins, CO.

Earlier publications on this approach include:

  • Crawford-Hines, S. and Anderson, C.W. (2000) Learning Expert Delineations in Biomedical Image Segmentation. In Proceedings of the Conference on Artificial Neural Networks In Engineering, ANNIE-2000, St. Louis, Missouri, November 5-8, 2000, pp. 657–662.
  • Crawford-Hines, S., and Anderson, C.W. (1998) Machine Learned Contours to Assist Boundary Tracing. In Proceedings of the IEEE Southwest Symposium on Image Analysis and Interpretation, Tucson, AZ, April, 1998.
  • Crawford-Hines, S., and Anderson, C.W. (1997) Neural Nets in Boundary Tracing Tasks. In Neural Networks for Signal Processing, VII, Proceedings of the 1997 IEEE Workshop, ed. by J. Principe, L. Giles, N. Morgan, and E. Wilson, pp. 207–215.
  • S. Crawford-Hines, and C. Anderson. Interactive Region Bounding with Neural Networks. Proceedings of NNACIP’94, the IEEE International Workshop on Neural Networks Applied to Control and Image Processing, 1994, pp. 58–61.

Converting Polygon Meshes to NURBS

Visible Productions, Inc., of Fort Collins, CO, produces 3-D human models that are recognized as some of the most accurate models in the world. Their models currently are based on meshes of 3-D triangles. Such meshes can be rendered as smooth surfaces by interpolating color values across a triangular mesh, but for a number of applications the smooth surface must be explicitly represented. Clients for Visible Productions’ models have asked for surfaces defined by NURBS (Non-Uniform Rational B-Splines). This project will develop and implement algorithms for transforming polygonal meshes into NURBS. This requires a time-intensive, interactive optimization process. We investigated the use of neural networks to by-pass a large part of the optimization process. Our recent projects have been funded by CASI and by an NSF SBIR grant. Material available from this work includes:

Terrain Modeling

The Forest Service often needs to visualize a given terrain at various stages of forest growth. This need, plus the fact that terrain elevation data is abundant while texture and color data of terrain is not, has led us to the following study. With the help of Denis Dean in the Department of Forest Sciences at CSU, we have trained neural networks to predict the color of a small area of terrain given only the elevation data:

Fast Calculation of Radiosity Form Factors

One of the most computationally expensive steps in radiosity is the calculation of the form factors between any two patches in a scene. Building a neural network approximation to the form factor calculation could drastically reduce the complexity of this calculation, making real-time calculation of form factors possible. This has been investigated by Charles Martin in

Leave a Comment