基于单帧图像的超分辨率算法

随着稀疏编码与压缩传感理论的逐步发展,如何应用于图像的超分辨率成为研究热点之一.基于示例学习的算法,提出了一种新的超分辨率算法,其特点在于只基于低分辨率图像本身,没有额外的样本库,运用自然图像的自相似性与冗余性,学习低分辨率图像块与高分辨率图像块之间的函数关系.为了从图像中获取更加全面的信息,采用Guided滤波、一阶导数和二阶导数2种方法来提取特征.此外,提出了一种新的字典学习算法R-KSVD,并且改进了后项处理过程.实验结果显示,提出的算法具有较好的超分辨率效果和稳定性.

Super-resolution algorithm based on a single low-resolution image

With the development of sparse coding and compressive sensing,image super-resolution reconstruction attracted extensive attentions.Based on the example-based algorithm,it was proposed a new super-resolution method.It exploited the relationship between the low image patches and the high image patches by the self-similarity of a natural image.The proposed method applied guided filter,the first-order and second-order derivatives to extract two kinds of features from the LR image,which was superior to using only one feature space.Besides,the effective dictionary was constructed by a novel algorithm called Relaxation K-SVD（R-KSVD）.Moreover,a new approach was proposed to estimate better HR residual image in the Back Projection.Experimental results demonstrated the superiority of the algorithm in both visual fidelity and numerical measures.