基于自类推的NSCT域单幅图像超分辨率重建  

NSCT Domain Single Image Super-resolution Reconstruction Based on Self Analogies

作　　者：程倩倩[1] 范新南[1] 李庆武[1] 

机构地区：[1]河海大学计算机与信息学院,常州213022

出　　处：《电子与信息学报》2011年第12期2881-2887,共7页Journal of Electronics & Information Technology

基　　金：国家自然科学基金(60972101;60872096);疏浚技术教育部工程研究中心开放基金(HDCN08002);江苏省社会发展科技项目资金(BS2007058)资助课题

摘　　要：单幅图像放大是一个病态问题。本文利用图像局部结构的自相似性和可传递性,结合非下采样Contourlet变换(NSCT)的优点,提出一种基于自类推的NSCT域单幅图像超分辨率重建方法。首先采用NSCT对源图像和退化图像进行多尺度、多方向分解,得到用于学习的低通子带对和各带通方向子带对,再利用图像自类推技术生成高分辨率的低通子带和各带通方向子带,最后进行NSCT重构得到超分辨率重建的图像。实验结果表明,该方法可以独立进行,摆脱一般方法对训练集合的依赖,并且较一般的图像类推算法速度大为加快,能产生更为合理的细节,视觉边缘更清晰,图像更逼真。Single-image zooming is an ill-posed problem. Using the self-similarity feature among local structure in an image which can be maintained in the scale space and the advantage of NonSubsampled Contourlet Transform （NSCT）, a single image super-resolution reconstruction algorithm based on image analogies in NSCT domain is proposed. Firstly, NSCT is performed on the original image and the degraded image at different scales and directions, thus low-pass subband pair and varieties of directional bandpass subband pairs are obtained. Then the high resolution low-pass subband and varieties of directional bandpass subband are generated by using image self-analogies. Finally, the super-resolution reconstructed image is obtained by transforming these subband coefficients back to the spatial domain. The experimental results show that the algorithm can be executed independently without any supposed outliers and it can compute much more sharply than general image analogies methods. It also can generate more reasonable details than general image analogies methods, thus the edges are much clearer and the image is more natural-looking.

关 键 词：图像处理 超分辨率 非下采样CONTOURLET变换 图像类推 

分 类 号：TN911.73[电子电信—通信与信息系统]