Model-based least squares optimal interpolation

The traditional approach to image interpolation is by synthesis using basis functions because of its computational simplicity and experience-proven quality of the result. We offer an alternative approach to designing the basis (interpolation kernels), using least-squares optimisation and image models that encompass the prior knowledge. In this paper we consider and derive a finite-support interpolation kernel based on a step-edge model and show that this results in a piece-wise cubic polynomial similar to Keys' cubic convolution. We offer an experimental comparison of the proposed kernel to a number of common methods and show that it performs similar to, or better than, the existing methods with similar extent of spatial support.