Interpolation of scientific image databases

This paper explores how recent convolutional neural network (CNN)-based techniques can be used to interpolate images inside scientific image databases. These databases are frequently used for the interactive visualization of large-scale simulations, where images correspond to samples of the parameter space (e.g., timesteps, isovalues, thresholds, etc.) and the visualization space (e.g., camera locations, clipping planes, etc.). These databases can be browsed post hoc along the sampling axis to emulate real-time interaction with large-scale datasets. However, the resulting databases are limited to their contained images, i.e., the sampling points. In this paper, we explore how efficiently and accurately CNN-based techniques can derive new images by interpolating database elements. We demonstrate on several real-world examples that the size of databases can be further reduced by dropping samples that can be interpolated post hoc with an acceptable error, which we measure qualitatively and quantitatively.