Unweighted: 7%	Weighted: 4%	Weighted: 3%	Weighted: 2%
Comp. Rate = 16.04	Comp. Rate = 27.85	Comp. Rate = 44.07	Comp. Rate = 47.22
Figure 1: Comparison of visualized images: ray-casting and isosurface rendering

Isosurface extraction and direct volume rendering are two of the most frequently used visualization techniques in volume visualization, and they have been applied effectively in several disciplines, including medical imaging, computational fluid dynamics, molecular modeling, meteorology, etc. Since volume data sets tend to be very large in many applications, considerable effort has been made to develop optimized algorithms for both methods.

When interactive real-time applications are developed with very large volume data, the use of compression is often inevitable. Usually, lossy compression techniques are employed since they achieve high compression rates. Lossy schemes, however, involve loss of information in reconstructed data, which may cause error-prone visual information. Although the methods based on the filters such as discrete cosine transform or Daubechies wavelets show excellent performances in both compression rates and image fidelity, they are often inappropriate for real-time applications that require fast decompression to random data access.

In many visualization applications, one often wishes to visualize some fixed set of coherent features that are concentrated in a few regions. For instance, isosurfaces with a few isovalues are built during isosurface extraction. In direct volume rendering, regions of voxels whose density values fall in the range of interest material are usually rendered. In case such to-be-examined features are known beforehand, it is possible to utilize the relevant information during compression to preserve them as precisely as possible. Lossy compression can be viewed as allocating a limited amount of resources to voxels, and it is the main goal of this work to wisely distribute them according to the specific types of visualization to be performed.

We present a technique that classifies voxels in accordance with the purpose of visualization, and that assigns weights to voxels properly. The associated weights are then combined with compression algorithms so that important features remain as correctly as possible after reconstruction. Contrary to the standard lossy compression schemes, in which information on features are uniformly lost regardless of the visualization to be performed, the new scheme enables one to focus more on important features that are more frequently used during the visualization. This approach is different from other enhancement techniques such as the classified vector quantization or the classification algorithm that attempt to improve the compression or classification quality based on spatial properties of images themselves.

Unweighted
	2%	3%	5%	7%
Weighted
	2%	3%	5%	7%
Figure 2: Comparison of visualized images: isosurface rendering

Unweighted
	2%	3%	5%	7%
Weighted
	2%	3%	5%	7%
Figure 3: Comparison of visualized images: ray-casting

Papers

C. Bajaj, I. Ihm, and S. Park, "Visualization-Specific Compression of Large Volume Data", Pacific Graphics 2001, Tokyo, Japan, 2001 (To appear). [pdf]

© Sanghun Park, CCV / TICAM, The University of Texas at Austin, August 30, 2001