在基于示例学习的单幅图像超分辨率(SISＲ)重建中，假设从低分辨率 (LR)到 高分辨率 (HR) 图像块的映射关系是一对一的，但同一 LR块会与多个 HR块对应，导致 了 LR与 HR块的匹配误差． 为解决 HR复原块的失配问题，文中首先导出了 LR块主题模 式的概率模型，引入信号的隐藏主题这一种新的观察信息． 然后提出了一种基于块主题差 异和上下文最大概率的结构感知复原机制，通过主题模式与邻域块内容的关联，形成 LR块的流形描述;在重构中通过自适应主题决策树选择和节点回归矩阵映射，从相似的 LR流形信号中准确区分和复原 HR信号． 主题模型优化实验结果表明，文中基于主题约束信 息的算法比未引入隐藏主题的决策树 SISR方法的峰值信噪比(PSNR)值提升了 0． 25dB; 在 5 种算法的对比实验中，相对于稀疏字典 SISR方法，文中方法的 PSNR值平均提升了 0． 92dB，表明引入隐藏的主题信息和主题流形结构辨识是可行的．

In the process of single image super-resolution reconstruction (SISR) based on learning from examples， the mapping relation was assumed one-to-one from a low-resolution (LR) input to a high-resolution (HR) image patch． But in fact，one LR patch may relate to many HR patches，and thus leads to matching errors． To solve the mismatch problem of restored patch，the probability model of LR patch topic pattern was derived to express new observation information for hidden topics in LR signals． Then a structure-aware recovery mechanism with topic differences and context maximum probability was proposed，and LR manifold description was formed by relating topic modes to LR neighbor contents． The HR signal was accurately distinguished and reconstructed from similar LR manifold signals via an adaptive selection of topic decision trees and regression matrix of nodes． The topic mo- del optimization experiment demonstrates that the peak signal-to-noise ratio (PSNR) of our topic constraint SISR method is improved by 0. 25dB compared to that of the decision tree based SISR algorithm without introducing hid- den topics． In the comparative experiment of five algorithms，the average PSNR value of our SISR approach is im- proved by 0. 92dB compared to that of the sparse dictionary based SISR method． So the introduced hidden topic in- formation and topic-manifold structure identification are feasible．