基于过完备字典的鲁棒性单幅图像超分辨率重建模型及算法  

Robust Single Image Super-resolution Reconstruction Model and Algorithm Based on Over-complete Dictionary

作　　者：徐国明[1,2] 薛模根[1,3] 崔怀超[3] 

机构地区：[1]合肥工业大学计算机与信息学院,合肥230009 [2]解放军陆军军官学院基础部,合肥230031 [3]解放军陆军军官学院信息工程系,合肥230031

出　　处：《计算机辅助设计与图形学学报》2012年第12期1599-1605,共7页Journal of Computer-Aided Design & Computer Graphics

基　　金：安徽省自然科学基金(1208085QF115)

摘　　要：针对单幅含噪图像的超分辨率重建问题,基于图像在过完备字典下的稀疏表示建立了超分辨率重建模型.该模型中低分辨率字典采用K-SVD算法直接训练,高分辨率字典则由高分辨率图像块与低分辨率字典下的同构的表示系数进行逼近求得;近似的高分辨率图像块通过高分辨率字典乘以表示系数得到,为使重建结果对噪声具有鲁棒性,利用基于稀疏表示的噪声图像恢复的方法由重叠的近似高分辨率图像块求得最终结果.实验结果表明,文中模型无论是主观视觉还是客观评价指标均取得了较好的效果,并验证了模型及算法的有效性.Aiming at the problem of super-resolution reconstruction for single noised image, in terms of sparse representation of over-complete dictionary, a super-resolution model is proposed. The K-SVD algorithm is used directly for learning the dictionary for low-resolution images. The dictionary for high-resolution images is got by optimizing the approximating error of the isomorphic sparse representation coefficients, which are got by learning the dictionary for low-resolution images. The representation coefficients are multiplied by the high-resolution dictionary to get the approximative high-resolution image patches. To make the reconstructed image robust to noise, the denoising method via sparse representation is used to get the final image from the overlapped approximative high- resolution image patches. The experimental results show that the proposed model obtains better outcome both in subjective visual effect and objective evaluation criteria, and demonstrates the effective of the model and algorithm.

关 键 词：超分辨率 单幅图像 稀疏表示 过完备字典 K—SVD 

分 类 号：TN911.73[电子电信—通信与信息系统]