结合非局部低秩先验的图像超分辨重建概率模型

Super-Resolution Reconstruction Using Probability Model Combined with Nonlocal Low-Rank Prior

现今图像成像技术日益普及,但受成像设备、成像环境以及在获取图像过程中外界噪声等因素的相互制约,在实际应用中很多图像成像分辨率较低,带来诸多问题.为此,提出一种有效的基于最大后验概率和非局部低秩先验的图像超分辨重建模型.首先,该模型采用连续图像序列作为数据输入,利用单幅图像内与连续图像间的相似性作为先验知识,提升相似图像块匹配度,消除图像细节丢失现象.然后,以最大后验概率框架建模,使用高斯分布和吉布斯分布拟合模型参数,提升模型泛化能力.通过相似块的奇异值估计待求块的奇异值,采用低秩截断抑制重建过程中引入的噪声.最后,利用图像的非局部自相似性和低秩性质,以非局部低秩约束正则化图像重建过程,添加图像的局部和全局信息来提升重建效果.在标准光流数据集、纽约大学和山东省千佛山医院提供的数据集上的实验结果表明,文中基于最大后验和非局部低秩先验的模型与传统插值算法、基于重建的优秀算法相比,在5组仿真实验中,其平均峰值信噪比提升6.3 dB,在保持图像纹理特征和恢复图像细节方面可取得更好的重建性能.

Nowadays imaging technologies have become more and more popular.However,due to the mutual constraints from imaging equipment,environment and other factors,such as external noise in the process of acquiring images,the image resolution is generally low in the practical applications,which causes many problems.In this paper,we propose an effective image super-resolution reconstruction model based on maximum a posterior probability(MAP)and nonlocal low-rank prior(NLP).Firstly,by inputting the continuous image sequence,the similarity inside the single image and among the image sequence is used as prior knowledge,in order to improve the matching degree of similar image patches and eliminate the loss of image details.Then,the reconstruction is modeled with MAP framework,where the parameters are fitted by Gaussian distribution and Gibbs distribution,respectively,for increasing the generalization capability.Furthermore,this model estimates the singular values of the desired patches by singular values of similar patches,and suppresses the noise by low-rank truncation.Finally,to exploit the nonlocal self-similarity and low-rank nature of images,NLP regularization is adopted to regularize the reconstruction process,which introduces the local and global image information to improve the reconstruction effect.The experimental results on the standard optical flow datasets and the datasets provided by New York University and Shandong Provincial Qianfoshan Hospital show that,this proposed model based on MAP and NLP is comparable to the traditional interpolation algorithms and the excellent reconstruction-based algorithms.This method increases the average peak signal-to-noise ratio by 6.3 dB in five simulation experiments and achieves better reconstruction performance in maintaining image texture features and restoring image details.