The conventional super-resolution algorithms on sparse representation reconstruct the high resolution image using one-stage high/low resolution dictionary pairs with inadequate detail information.In order to recover detail information as much as possible,two-stage dictionaries are explored in this paper.Then we train jointly multiple-frequency-band dictionaries consisting of low frequency (LF) dictionaries,middle frequency (MF) dictionaries and high frequency (HF) dictionaries,and simultaneously exploit the prediction relation of LF component,MF component and HF component to recover middle and high frequency information.Considering that there are many repetitive structures in the natural image,nonlocal self-similarity information is combined properly with iterative back-projection procedure to post-process the image.Experimental results demonstrate that the proposed algorithm has remarkable improvement in peak signal-to-noise ratio,structural similarity and visual quality compared with the other learning-based algorithms.%常规基于稀疏表示的超分辨率算法使用一级高低分辨字典重构图像,恢复细节信息不充分.本文利用两级字典恢复尽可能多的细节信息;然后构造联合低频字典、中频字典、高频字典的分频带字典,利用图像低频、中频、高频三者之间的预测关系,恢复图像中的高频信息.利用图像的非局部相似性,将其与迭代反向投影算法相结合,进行图像的后处理.实验结果表明,与其他几种基于学习的算法相比,本算法无论是在峰值信噪比、结构相似性指标,还是视觉效果上都有显著的提高.
展开▼