图像及视频盲超分辨率重建及其关键技术研究

High resolution images and videos are often required in many areas such as medical, remote sensing, military, intelligent monitoring etc. How to improve the resolution of images and videos acquired by the current imaging devices becomes an urgent problem. Super resolution reconstruction technology is an efficient way to solve this problem with low cost and high realizability. At present, the research on image and video blind super resolution reconstruction is still on elementary stage. Some key technologies have not been well resolved, which need further studies. The main innovative points are expressed as follows: (1) In order to extend the applicable range of the registration algorithm, non-rigid motion registration methods are researched. Motion models with five parameters, six parameters and ten parameters are proposed respectively to registrate affine transformation and projective transformation, and a new force is used to estimate elastic transformation; (2)In order to improve the quality of super resolution reconstructed image and video,Gaussian,Motion and defocus point spread function estimation methods are proposed, and point spread function in the case of multiple blur coexisted estimation method is researched; (3) in order to restrain the ghost phenomenon appeared in reconstructed images and videos, a new regularized model is proposed,in which the motion registration error and degrade function error are added; (4) In order to improve the applicablity of our algorithm,the blind super resolution reconstruction methods of videos with large global motion and large local motion are proposed respectively.

在医学、遥感、军事、智能监控等许多领域中,需要获得高分辨率的图像及视频，如何提高现有成像设备所得到的图像及视频的分辨率成为迫切需要解决的问题，超分辨率重建技术是解决该问题成本低、可实现性高的有效方法。目前，图像及视频盲超分辨率重建技术还处于初级研究阶段，部分关键技术尚未得到较好解决，亟待更深入地研究。本项目的主要创新点有：(1)对非刚体运动配准算法进行研究，提出五参数、六参数和十参数运动模型分别对仿射变换及投影变换进行配准，使用一个新的作用力对弹性变换进行配准，扩大配准算法的适用范围；（2）提出高斯、运动、散焦点扩展函数的估计方法，对多种模糊共存时的点扩展函数进行研究，提高超分辨率重建图像及视频的质量；(3)提出新的正则化模型，引入运动配准误差及降晰函数估计误差，抑制重建图像及视频中的"鬼影现象"；（4）提出帧间存在较大全局运动及较大局部运动时的视频盲超分辨率重建方法，提高算法的实用性。

本项目针对不同情况的图像及视频，对其盲超分辨率重建模型及部分关键技术进行研究，主要包括：（1）在图像盲超分辨率重建中，分别提出了单幅及多幅图像盲超分辨率重建模型；（2）在视频盲超分辨率重建技术中，分别提出了视频单帧盲超分辨率重建模型，以及基于滑动窗口的视频多帧盲超分辨率分段重建模型；（3）在运动配准算法中，针对存在不同运动情况时的多幅图像，通过高斯金字塔模型及迭代求解提高配准算法的精度和适用范围，减少“误配准”现象；（4）在降晰函数估计中，分别对存在高斯、运动、散焦模糊的图像进行盲超分辨率重建，使用误差-参数分析方法及频率域方法估计点扩展函数，抑制“鬼影”现象；（5）在超分辨率重建算法中，采用改进的迭代反投影算法，充分利用估计到的相关信息进行重建，并进行正则化处理，提高重建图像及视频的质量；（6）提出采用视频多帧重建方法对帧间存在较大全局运动和较小局部运动的视频进行重建，采用视频单帧重建方法对帧间存在较大局部运动的视频进行重建；（7）提出单幅含噪图像及多幅含噪图像的超分辨率重建方法，减少了噪声对重建重量的影响；（8）设计制作图像盲超分辨率重建系统、以及视频盲超分辨率重建系统，为技术推广应用奠定了坚实的基础。