基于近邻嵌入逐级放大的图像超分辨率重建

Super-Resolution Image Reconstruction Based on Stepwise Magnification of Neighbor Embedding

针对基于近邻嵌入的图像超分辨率重建,提出带约束的逐级放大策略来提高近邻保持率,改进重建效果,并对各级放大的图像用迭代反向投影约束进行修正,减少学习过程中可能出现的误差,保证每一级的解向着正确的方向演化.此外,为充分利用测试图像本身的信息,将由测试图像得到的在线训练集与由训练图像数据库得到的离线训练集串联,构成联合训练集,进一步改进算法的性能.实验表明,与现有的一些算法相比,文中算法无论在视觉效果还是客观评价上都获得了更好的结果.

Proposed in this paper is a constrained stepwise magnification strategy for the super-resolution image reconstruction based on the neighbor embedding, which is used to increase the neighborhood-preserving rate and improve the reconstruction effect. Then, the iterative back-projection constraint is used to modify the magnified image in each step, which decreases the errors that may occur during the learning procedure and makes the solution of each step evolve in a correct direction. Moreover, in order to take full advantage of the information of the test image, a joint training set is constructed by concatenating the on-line training set learned from the test image and the offline training set learned from the training image database, and is used to further improve the algorithm performance. Experimental results show that, as compared with some existing algorithms, the proposed algorithm helps to obtain better visual effect and more objective evaluation results.