¹Georgia Institute of Technology, Atlanta, GA, USA
²Lawrence Livermore National Laboratory, Livermore, CA, USA

---

Abstract—Many scientific, imaging, and geospatial applications produce large high-precision scalar fields sampled on a regular grid. Lossless compression of such data is commonly done using predictive coding, in which weighted combinations of previously coded samples known to both encoder and decoder are used to predict subsequent nearby samples. In hierarchical, incremental, or selective transmission, the spatial pattern of the known neighbors is often irregular and varies from one sample to the next, which precludes prediction based on a single stencil and fixed set of weights. To handle such situations and make the best use of available neighboring samples, we propose a local spectral predictor that offers optimal prediction by tailoring the weights to each configuration of known nearby samples. These weights may be precomputed and stored in a small lookup table. We show through several applications that predictive coding using our spectral predictor improves compression for various sources of high-precision data.

Index Terms—interpolation, prediction, compression, spectral basis, discrete cosine transform, irregular stencils

[PDF]

Cite: Lorenzo Ibarria, Peter Lindstrom, Jarek Rossignac, "Spectral Interpolation on 3×3 Stencils for Prediction and Compression," Journal of Computers vol. 2, no. 8, pp. 53-63 , 2007.