Super-Resolution Reconstruction Based on Kernel Regression Method

Abstract

Kernel regression method is one of the non-parametric nonlinear regression estimation methods. The univariate regression function is estimated by one-dimensional kernel function. Based on Taylor expansion, the value of regression coefficient is obtained by optimization method. Using the same method, two-dimensional kernel regression is studied. Finally, the adaptive control kernel is summarized. On the basis of the non-parametric estimation kernel regression model, the two-dimensional kernel regression function is extended to three-dimensional and each pixel in the video sequence is represented as a three-dimensional Taylor expansion. The displayed coefficients are obtained by the locally weighted least squares method, and the weights of the kernel regression are used to capture the spatiotemporal local motion information, avoiding the explicit sub-pixel precision motion estimation. Experiments in the standard test video database show that the algorithm has better reconstruction effect and larger scope of application, and can be used for videos with local and complex motion. The 3D-SKR super-resolution reconstruction algorithm solves the sub-pixel accuracy registration problem of the previous super-resolution reconstruction algorithms and extends the application range of sequence image super-resolution reconstruction to any sequence image, which is of great significance. The new algorithm framework proposed in this paper improves the robustness of the original algorithm and eliminates the influence of outliers on the reconstruction results. The algorithm proposed in this paper can adaptively capture the structural information of the image, has a good suppression effect on noise, and can effectively eliminate the artifacts caused by inaccurate motion estimation.