基于自相似性约束的视频稀疏超分辨率重建

Video Super-resolution Based on Sparse Representation with Non-local Self-similarity Regularization

通过研究帧间自相似性对图像重建的影响,提出一种自相似性约束的单视频稀疏超分辨率重建算法,以达到保持图像局部结构完整性的同时有效去噪的目的。该算法运用主成分分析PCA训练出适应图像不同局部结构的分类词典;通过帧间光流场的粗略运动估计和帧内帧间的精确块匹配,搜索自相似信息,运用非局部均值NLM滤波,并以此约束稀疏模型。仿真实验表明,提出的算法无论是客观指标,还是主观视觉上都超过了进行比较的几种分辨率提高算法。

By studying the inter-frame self-similarity on the image reconstruction, a method for single video super resolution（SR） based on sparse repre- sentation with self-similarity constraints is proposed in this paper, aimed to maintain structural integrity of local image while de-nose effectively. In this meth- od, the skill of principal component analysis （PCA） is used to learn dictionary of several classes from which different local structure of iraage can adaptively select a sub-dictionary as a sparse domain; the serf-similarity redundant information, which is used for non-local means （NLM） filtering, can be gained through firstly coarse inter-frame motion estimation in the optical flow field, then accurate inter/intra block matching, and to constrain the sparse reconstruc- tion model. Extensive experimental comparisons with sate-of-the-art SR validated the generality and effectiveness of the proposed mathod.