多帧超分辨率重建中的图像插值

为了实现多帧超分辨率重建,必须从图像序列中提取子像素信息,因此要求图像的配准精确到子像素级。本文提出了一种超分辨率重建中运动矩阵的构造方法。为了提高矩阵的构造精度,该方法引入图像插值,用性能好但计算量较小的三次插值算子实现插值运算,并比较了三种常用插值核的性能。实验结果表明,选择适当的插值函数不仅不会明显增加计算量,而且可以显著提高矩阵的构造精度,从而大大提高多帧超分辨率重建图像的质量。

Image Interpolation in Superresolution from Multiple Frames

To realize superresolution from multiple frames, subpixel information content should be extracted from the sequence. This requires that each frame be registered accurately with respect to some reference to within subpixel accuracy That is to say, in the classical superresolution restoration model, the accurate construction of warp matrix is of great importance. In this paper, a method of constructing the motion matrix is proposed. To improve the accuracy of construction, image interpolation is introduced. The performances of three commonly utilized interpolations are tested: nearest, linear and cubic interpolation. Experimental results show that much better performance can be achieved in superresolution from multiple frames while introducing cubic interpolation, moreover, the computational cost is not obviously higher than those of the other two interpolations.