图像超分辨率重建中的运动模型选择

On Motion Model Selection in Image Super-resolution Reconstruction

传统的超分辨率方法通常假定一个低分辨率图像序列所服从的固定的运动模型,但是这种假设并不能覆盖实际应用中出现的各种运动情况,从而严重影响图像的重建质量。本文提出一个"虚拟像机"的概念,将序列中的每一幅低分辨率图像映射到一张高分辨率参考图像的网格上,并引入运动模型选择来同时估计出两视点几何关系和相关参数。文中给出了应用GRIC(Geometric Robust Information Criterion)准则确定两视点几何关系的方法。

Conventional SR algorithms usually assume that a fixed motion model is followed by all LR image sequences, which is, however, an impractical assumption for many real applications, and therefore affects the reconstruction quality heavily. This paper proposes a concept, Virtual Camera, to map each LR image onto a high-resolution reference image grid and introduces motion model selection to simultaneously estimate the epipolar geometry, and the relevant parameters. The method using the GRIC （ Geometric Robust Information Criterion） criterion to determine the epipolar geometry is presented in this paper.