基于正则化稀疏表示的图像超分辨率算法  

Super-resolution image reconstruction algorithm via regularized sparse representation

作　　者：朱波[1,2] 李华[3] 高伟[1] 宋宗玺[1] 

机构地区：[1]中国科学院西安光学精密机械研究所,陕西西安710119 [2]西安交通大学西安光学精密机械研究所空间视觉联合实验室,陕西西安710049 [3]商洛学院,陕西商洛726000

出　　处：《光电子．激光》2013年第10期2024-2030,共7页Journal of Optoelectronics·Laser

基　　金：国家自然科学基金(90920301);陕西省教育厅科研计划(2013JK1057)资助项目

摘　　要：为了从单幅低分辨率(LR)图像恢复出高分辨率(HR)图像,提出了一种应用正则化稀疏表示和基于机器学习的超分辨率(SR)图像恢复算法。构造了一种基于稀疏表示的SR凸变模型,为了提高恢复效果,针对模型提出了两种稀疏正则化约束条件,一是将分类效果更好的图表拉普拉斯作为正则化约束条件,从而找到与输入LR图像块在结构上最接近的学习样本;另一种是针对冗余的学习样本进行约束,保证了图像边缘的锐利。将输入的每一块LR图像应用正则化稀疏表示,经过学习得到与之对应的HR图像块,最终得到整幅HR图像。试验结果表明,算法恢复出的HR图像峰值信噪比(PSNR)值较双三次插值算法最高提升约2dB,主观目视清晰、边缘锐利。In order to produce a higher resolution image from a low resolution one, a machine learning su- per-resolution image reconstruction method via regularized sparse representation is presented. A convex variational model is proposed for image super-resolution with sparse representation regularization. We further introduce two adaptive regularization terms into the sparse representation framework to improve the processing effect. Firstly, the Graph Laplacian based image clustering model which takes the local manifold structure into account to a given patch is selected to regularize the image local structures. See- ondly, the image nonlocal self-similarity is introduced as another regularization term,which improves the quality of reconstructed images. Therefore,the sparse representation of a low-resolution image patch can be applied with the high-resolution image patch dictionary to generate a high-resolution image patch,and then, the high-resolution image is obtained. Experimental results proved that the proposed algorithm out performs some state-of-the-art super-resolution methods both in peak signal-to-noise ratio （PSNR） measure and visual quality.

关 键 词：超分辨率(SR) 稀疏表示 图像分类 正则化 图表拉普拉斯 

分 类 号：TN911.73[电子电信—通信与信息系统]