方向字典子图的初始邻域嵌入重构

邻域嵌入超分辨率重构算法在空间邻域选取过程中,细节特征易被大幅度特征分量淹没,为此,提出了基于方向字典子图的初始邻域嵌入重构算法.对输入图像及邻域利用方向字典进行稀疏分解,从大、小幅值表示系数中分别重构大、小幅度特征子图,保护邻域计算中的小幅度特征；同时,为降低多子图重构的运算量,通过随机森林机制,将输入图像在分类树森林中对应叶子节点图像子库的并集作为初始邻域,减小实际参与运算的图像库大小.实验结果表明,相对于邻域嵌入超分辨率算法,基于方向字典子图的初始邻域嵌入重构的峰值信噪比值平均提升了1.0959 dB,有效改善了重构效果；重构时间仅为邻域嵌入超分辨率的13.3%,降低了重构复杂度.

Initial Neighbor Embedding Super Resolution Based on Sub-Image of Directional Dictionary

In process of choosing spatial neighborhood in neighbor embedding based super resolution ( NESR) , the detail features may be covered by large value ones, the Initial neighbor embedding super resolution based on sub-image of directional dictionary ( DD-NESR) was proposed. It decomposes input image and neighborhood by directional sparse dictionary, which is utilized for representation of large and tiny value to reconstruct their sub-images respectively, so as to protect detail features in neighborhood calculation. In order to reduce reconstruction complexity, the author uses random forests ( RF) mecha-nism to obtain node sub-images of input in classification tree forest, and takes union of sub-images for ini-tial neighborhood, to reduce size of calculating database. It is shown that, comparing to NESR, DD-NESR improves PSNR in 1. 095 9 dB averagely, the method promotes reconstruction quality effectively;its reconstruction time is 13. 3% of NESR, with lower complexity.