基于块旋转和清晰度的图像超分辨率重建算法  

An Image Super Resolution Reconstruction Algorithm Based on Patch Rotation and Sharpness

作　　者：尧潞阳 解凯[1] 李桐[1] 

机构地区：[1]北京印刷学院信息工程学院,北京102600

出　　处：《计算机技术与发展》2017年第1期70-74,共5页Computer Technology and Development

基　　金：国家自然科学基金资助项目(61472036)

摘　　要：针对传统基于学习的超分辨率重建算法训练时间过长,且对训练库依赖性大的不足,提出一种结合块旋转和清晰度的超分辨率重建方法。该方法引入了一种新的分类机制。为增加训练样本块的多样性,但又不增加计算复杂度,将样本块进行一定角度的旋转,然后引入块清晰度(Sharpness Measure,SM)对训练样本进行分类。对于块清晰度较高的纹理、角以及边缘块,利用分类好的对应样本库进行自相似性重建,而清晰度较低的块,则直接使用插值放大进行重建。在搜索匹配过程中改用Fast Library for Approximate Nearest Neighbors(FLANN)替代传统的Approximate Nearest Neighbors(ANN)搜索,提高了重建效率。最终利用迭代反投影算法和局部约束进行优化。实验结果表明,该算法既可以较大幅度减少计算的复杂度,也能够获得较好的视觉效果。To address the shortcomings of long-time training and relying on the additional training databases in conventional example- based super-resolution algorithm, a super-resolution image reconstruction algorithm based on patch rotation and sharpness is proposed, which introduces a new classification mechanism. To increase the diversity of training sample patches but not the computational complexi- ty ,they are rotated by a certain angle and then introduced the patch Sharpness Measure （SM） to classify the training samples. For patches of high SM, such as textures, comers and edges, the self-similarity reconstruction is carried on by classified samples. For patches of low SM, the interpolation is used directly to enlarge the image for reconstruction. During the searching and matching process, the Fast Library for Approximate Nearest Neighbors （FLANN） to replace the traditional Approximate Nearest Neighbors （ANN） increases the recon- struction efficiency. Finally, iterative back projection and local constraint are used for optimization. Experimental results validate that the algorithm not only can reduce the computational complexity effectively but also achieve better visual effects.

关 键 词：超分辨率 多尺度自相似性 块旋转 清晰度 迭代反投影 局部约束 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]