仿射运动模型下的图像盲超分辨率重建算法

Blind Super-Resolution Reconstruction Algorithm under Affine Motion Model

研究利用帧间存在仿射运动的低分辨率图像序列重建出更高光学分辨率图像的盲超分辨率(BSR)问题.首先给出一种基于特征向量的模糊核零空间矩阵构造方法.将模糊的零子空间约束作为一项规整化泛函,提出一种非参数化模糊辨识、运动估计和图像重建三重耦合问题的联合迭代算法.该算法采用一个二层优化策略:先将三重耦合的BSR问题分解为关于模糊的二次型和关于运动参数与图像的非线性最小二乘(NLS)问题,再采用Gauss-Newton方法求解该NLS问题.仿真实验结果表明,文中提出的仿射变换下的BSR算法能对图像空间移变退化过程进行更为精确的建模,比纯平移BSR算法有更强的局部纹理恢复能力.最后通过真实车牌图像序列展示该算法的适用性.

An approach to the blind super-resolution （BSR） problem is proposed which yields a higher optical resolution image from a low-resolution （LR） image sequence with affine inter-frame motion. Firstly, an eigenvector-based method for constructing the null space of blurs is presented. It is used as the regularization constraint of the optimization procedure. Then, the iterative algorithm is developed for the triple-coupled problem. The proposed algorithm adopts a two-layer optimization strategy： in the first layer, the triple-coupled BSR problem is reduced to a quadratic form with respect to the blurs, and an nonlinear least squares （NLS） problem of the motion and the high-resolution （HR） image; in the second layer, the NLS problem is solved using a Gauss-Newton based method. The experimental results on synthetic data illustrate that the proposed BSR algorithm for affine transform, has better performance in terms of modeling the space-variant degradation process as well as restoring the local textures compared with the BSR algorithm for pure translation. Finally, the applicability of the proposed algorithm is demonstrated using real license plate images.