基于多尺度块搜索的单幅图像超分辨率重建

Single-image Super-resolution Reconstruction Based on Multiple Scales Search

针对基于学习的超分辨率重建算法重建结果不自然,且重建速度较慢等问题,提出一种在多级下采样图像中搜索相似块并自适应加权的重建方法。对学习到的高分辨率块进行修正,并根据相邻块重叠部分的相似度对上述高分辨率块进行加权,从而降低重建后图像的模糊度和锯齿现象。采用随机块搜索方法对相似块进行查找,相对于树结构的查找方法大幅降低了重建时间。实验结果证明,该方法能够在无任何先验信息的条件下,快速地对单幅图像进行超分辨率重建,并且从有参和无参的图像质量评价两方面,证明重建出的图像质量也有所提高。

Previous super-resolution reconstruction methods based on learning fail to obtain no artificial result and cost much time.Aiming at the problem,the reconstruction method by searching similar patches in multiple scales and weighing them adaptively is proposed.Through modifying the learned high resolution block and taking into consideration of the mutual influence among surrounding pixels,the reconstruction ambiguity and saw tooth on the edge of image are reduced.Furthermore,this paper introduces randomized patch-match for searching similar patches which costs less time than the methods using tree structures.Experimental results show that the proposed algorithm can reconstruct high resolution image rapidly for a single low resolution image without any priori information.The proposed method also improves and validates the quality of super-resolution reconstruction with the help of the reference image based quality assessment and the blind image quality assessment.