Neighbor embedding based super-resolution algorithm through edge detection and feature selection

Assuming that the local geometry of low-resolution image patches is similar to that of the high-resolution counterparts, neighbor embedding based super-resolution methods learn a high-resolution image from one or more low-resolution input images by embedding its patches optimally with training ones. However, their performance suffers from inappropriate choices of features, neighborhood sizes and training patches. To address the issues, we propose an extended Neighbor embedding based super-resolution through edge detection and Feature Selection (henceforth NeedFS). Three major contributions of NeedFS are: (1) A new combination of features are proposed, which preserve edges and smoothen color regions better; (2) the training patches are learned discriminately with different neighborhood sizes based on edge detection; (3) only those edge training patches are bootstrapped to provide extra useful information with least redundancy. Experiments show that NeedFS performs better in both quantitative and qualitative evaluation. NeedFS is also robust even with a very limited training set and thus is promising for real applications.