Covariance‐Based Interpolation
Chi-Wah Kok, Wing-Shan Tam · 2018
The formulation and accuracy of the linear prediction involves the statistical model of a natural image and the minimization of the prediction error by considering the “minimum mean squares error” (MMSE). This chapter discusses the second-order statistical model of natural image and the basic of MMSE optimization-based image interpolation method. New Edge-Directed Interpolation (NEDI) method assumes the image model assumptions: natural image is a locally stationary second-order Gaussian process; and low-resolution and high-resolution images have similar second-order statistics in corresponding local patches. One of the reasons leading to the covariance mismatch of the NEDI is the application of fixed window structure, both size and shape. As a result, one of the remedies is to modify the window to enclose the major feature only, such as to ensure that the low-resolution image and high-resolution image have similar covariance structures.