基于稀疏编码和各向异性滤波的超分辨率重建  

Super-resolution reconstruction based on sparse coding and anisotropic filtering

作　　者：伏伶丽 任超[1] 何小海[1] 吴晓红[1] Fu Lingli;Ren Chao;He Xiaohai;Wu Xiaohong(College of Electronics and Information Engineering,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学电子信息学院

出　　处：《信息技术与网络安全》2020年第3期23-28,共6页Information Technology and Network Security

基　　金：国家自然科学基金(61801316)

摘　　要：针对基于非局部稀疏自相似性的超分辨率重建方法存在的图像边缘保持性能不佳的问题,提出了一种基于稀疏编码和各向异性滤波的超分辨率重建算法。该算法利用卷积神经网络和各向异性引导滤波训练了一个各向异性特征模型,然后利用该特征模型构建一个局部的结构先验,以和非局部稀疏先验形成互补,从而提高算法的边缘保持能力。该算法训练后的模式使用通用测试集进行测试,测试结果表明算法SR性能较好,能很好地保持边缘细节,提供视觉效果更好的图像。To solve the problem of poor edge-preserving performance in the non-local sparse-based super-resolution reconstruction method,the algorithm based on sparse coding and anisotropic filtering is proposed.Specifically,an anisotropic feature model is trained by the convolution neural network and anisotropic guided filtering,and then a local structure prior is constructed by using the feature model to complement the non-local sparse prior,and thus improves the edge preserving ability of the algorithm.The trained model of this algorithm is tested with the general test data.The experimental results demonstrate that the proposed algorithm effectively improves the edge-preserving performance,and obtains higher objective evaluation indices and images with better visual effect.

关 键 词：稀疏编码 各向异性引导滤波 超分辨重建 卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]