high performance computing on graphics processing units: hgpu.org

Application of geometric transformation to images requires an interpolation step. When applied to image rotation, the presently most efficient GPU implementation for the cubic spline image interpolation still cost about 10 times as much as linear interpolation. This implementation involves two steps: a prefilter step performs a two-pass forward-backward recursive filter, then a cubic polynomial interpolation step is implemented thanks to a cascade of linear interpolations. This paper proposes a simpler and faster implementation for the prefilter – which is the most time consuming – in terms of a direct convolution. The overall cost for our cubic B-spline interpolation algorithm then reduces to only twice the cost of linear interpolation.