基于字典学习与结构自相似性的码本映射超分辨率算法

摘要

The spatial resolution of the image is limited by many factors including the environment of the imaging, the manufacturing technology of the hardware and the cost. A codebook mapping based single im-age super resolution (SR) method via dictionary learning and structural self-similarity is proposed to pro-mote the spatial resolution of the image. Firstly, the images in the training set are used to construct two codebooks corresponding to the low and the high frequency components of image patches, and the two codebooks are taken as the training samples to learn two dictionaries respectively. Then, the similar image patches of each input image patch in the initial reconstructed image are used to construct the nonlocal con-straint. Finally, the sparse representation coefficient of the input image patch is obtained by solving anl0 norm minimization problem with the nonlocal constraint, and the high resolution image patch is then recon-structed by using the low and the high dictionaries. The input image patch is represented by the dictionaries, rather than the codebooks, therefore the computational efficiency is increasing. Similar image patches are used to construct the regularization constraint, consequently the quality of the reconstructed image is im-proved. Experimental results demonstrate that the proposed method achieves better reconstructed results compared with LLE, ScSR and NARM methods.%图像的空间分辨率受成像环境、硬件制造水平和成本等多方面因素的影响,存在一定的局限性。为了提高图像的空间分辨率,提出一种基于字典学习与结构自相似性的码本映射超分辨率算法。首先利用训练集构建与图像高低频分量对应的高低频码本,将高低频码本作为训练样本获取高低频字典；然后在初始重建图像中搜索目标图像块的相似图像块,利用相似图像块构建非局部约束项；最后通过求解含有非局部约束项的 l0范数最小化问题获取目标图像块的稀疏表示系数,并利用高低频字典重建高分辨率图像块。该算法利用高低频字典表示目标图像块,而不是直接采用高低频码本,提高了算法的运算效率；利用相似图像块构建正则化约束项,提高了重建图像的质量。实验结果表明,与LLE, ScSR和NARM等算法相比,文中算法取得的超分辨率重建效果更好。