Super resolution with edge-constrained motion estimation

Yue Zhuo, Jiaying Liu, Mading Li, Zongming Guo

Abstract: Motion estimation is a critical step for most reconstruction-based super resolution methods. However, accurate motion estimation is difficult, and the unavoidable error degrades performance of super resolution rapidly. In this paper, we present a robust way to perform super resolution by improving motion estimation. Starting with feature points matching, we compute the local motion parameter of feature point correspondences by using the weighted Lucas-Kanade algorithm. Then accurate motion field is estimated by support region search, which refers to edge information and considers discontinuities of motion boundary and consistency of motion field. Experimental results validate the efficacy of each step in the proposed algorithm and show that it produces super resolved images with higher quality.