改进的基于邻域嵌入的图像超分辨率重构

Improved image super-resolution reconstruction based neighbor embedding

为了提高传统的基于邻域嵌入的图像超分辨率重构算法的时间效率,采用了一种利用图像块方向信息进行邻域选择和训练集分类的新方法。该方法首先利用图像块方向的不同对训练集进行分类,然后在分类后的子训练集中选择与待重构图像块的方向相似的图像块作为邻域进行重构,并对重构结果进行迭代反投影全局后处理,进一步提高重构质量,最后对改进方法进行数值实验验证。结果表明,该方法不仅把超分辨率重构的时间效率提高了10倍以上,而且重构质量也得到了改善,具有较好的实际应用价值。

In order to improve the time-efficiency of traditional super-resolution reconstruction based on neighbor embedding, a new method was proposed using direction information of image patches to choose neighborhood and classify the training set. Firstly, the training set was classified through the differences of patches directions. Secondly, the neighborhood used to reconstruct was chosen in the sub-sets by selecting training patches with the similar direction, and then the iterative back-projection was applied during the reconstruction to further enhance the super-resolution image quality. Finally, numerical experiments were conducted to verify the new method. The results show that the proposed algorithm increases time-efficiency more than 10 times and super-resolution performance is improved. The new method has better practical value.