关键词: 图像插值, 几何对偶性, 普通最小二乘, 非局部均方, 加权最小二乘, 核岭回归

Abstract: In order to protect the local structure of images better and improve the robustness of image interpolation algorithm, an improved image interpolation algorithm is proposed. It is based on a model that combines ordinary least squares based geometric duality and weighted least squares based non-local mean squares to form the steady-state region. Firstly, coefficients of weighted least square are estimated by non-local mean, and then are corrected by kernel ridge regression. Taking into account the bias of kernel ridge regression, ordinary least square is introduced as a regular term. The regularization parameters are adjusted adaptively. Compared with the algorithm with ordinary least squares or weighted least squares, the algorithm suppresses the interpolation artifacts effectively, and the results of image interpolation have higher peak signal to noise ratio and structural similarity.

Key words: image interpolation, geometric duality, ordinary least square, non-local mean square, weighted least square, kernel ridge regression