Efficient image interpolation by associating 2nd order local structure and data-adaptive kernel regression

Image interpolation is still a widely studied issue for rescaling a low-resolution image to a high resolution image. This paper tends to modify the data-dependent steering kernel regression image interpolation in order to reduce the computational cost in point-wise determination of data-dependent or nonlinear filter coefficients. Instead solving a kernel-based weighting least mean squared minimization a novel example-based matching approach is introduced and the problem is turned into the nearest neighbor search problem. Through conducted experiments applied to several images the proposed method is verified to reduce the computational time about 50% compared with the steering kernel regression algorithm while almost maintaining image quality.