IMAGE INTERPOLATION ON THE CPU AND GPU USING LINE RUN SEQUENCES

Abstract

Abstract. This paper describes an efficient implementation of an image interpolation algorithm based on inverse distance weighting (IDW). The time-consuming search for support pixels bordering the voids to be filled is facilitated through gapless sweeps of different directions over the image. The scanlines needed for the sweeps are constructed from a path prototype per orientation whose regular substructures get reused and shifted to produce aligned duplicates covering the entire input bitmap. The line set is followed concurrently to detect existing samples around nodata patches and compute the distance to the pixels to be newly set. Since the algorithm relies on integer line rasterization only and does not need auxiliary data structures beyond the output image and weight aggregation bitmap for intensity normalization, it will run on multi-core central and graphics processing units (CPUs and GPUs). Also, occluded support pixels of non-convex void patches are ignored, and over- or undersampling close-by and distant valid neighbors is compensated. Runtime and accuracy compared to generated IDW ground truth get evaluated for the CPU and GPU implementation of the algorithm on single-channel and multispectral bitmaps of various filling degrees.