基于回归与图像自相似的单幅图像超分辨率

发布时间：2019-05-08 17:25

【摘要】：超分辨率重建是利用单幅或多幅带有亚像素位移的低分辨率图像借助图像模型和降质先验等进行超分辨率的重构进而得到高分辨率图像,并在众多的图像应用领域中扮演着重要的角色,如:遥感、军事、医学成像、公共安全等等。超分辨率重建在数学上是一个不适定的逆问题,对其研究一直没间断过。目前超分辨率的重建主要有基于插值、基于重构和基于学习三种方法。本文围绕基于图像自相似与回归学习映射关系的单幅图像超分辨率方法进行展开。本文主要研究内容与贡献包括:(1)提出一种利用图像块自相似性和对非线性映射拟合较好的支持向量回归模型的单幅超分辨率算法。该方法不使用外部图像训练库,而是通过建立图像金字塔得到训练库,并假设低分辨率相似块和其对应的高分辨率图像块的中心像素之间的映射关系相同,利用支持向量回归模型学习映射关系。最后对其进行了仿真实验及相关分析,并与几个经典超分辨率方法进行比较,实验结果表明本文设计的算法能够很好的实现图像的超分辨率重建,特别是对纹理结构相似度高的图像具有更好的重构效果。(2)针对支持向量回归在此方法中的不足之处又提出了一种基于直接稀疏核回归的单幅图像超分辨率方法,寻找“更优”的支持向量,使得相似块的权重分配更接近实际情况。实验结果显示超分辨率效果确实有所提升。
[Abstract]:Super-resolution reconstruction is using single or multiple low-resolution images with sub-pixel displacement to obtain high-resolution images by super-resolution reconstruction by means of image models and degenerated priors. It plays an important role in many image applications, such as remote sensing, military, medical imaging, public safety and so on. Super-resolution reconstruction is an ill-posed inverse problem in mathematics. At present, there are three methods of super-resolution reconstruction: interpolation-based, reconstruction-based and learning-based. This paper focuses on a single image super-resolution method based on the relationship between image self-similarity and regression learning mapping. The main contents and contributions of this paper are as follows: (1) A single-frame super-resolution algorithm based on self-similarity of image blocks and better fitting of support vector regression model with non-linear mapping is proposed in this paper. This method does not use the external image training library, but builds the image pyramid to get the training library, and assumes that the mapping relationship between the low resolution similar block and the corresponding high resolution image block is the same as the central pixel of the corresponding high resolution image block. Support vector regression model is used to learn mapping relations. Finally, the simulation experiment and correlation analysis are carried out, and compared with several classical super-resolution methods. The experimental results show that the algorithm designed in this paper can realize the super-resolution reconstruction of the image very well. Especially, the image with high texture similarity has better reconstruction effect. (2) aiming at the deficiency of support vector regression in this method, a single image super-resolution method based on direct sparse kernel regression is proposed. Looking for "better" support vectors makes the weight assignment of similar blocks closer to the actual situation. The experimental results show that the super-resolution effect has indeed improved.
【学位授予单位】：天津大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】