Efficient Parallelization of Motion Estimation for Super-Resolution

This paper presents an efficient parallelization of the Motion Estimation procedure, one of the core parts of Super Resolution techniques. The algorithm considered is the basic version of Block Matching Super Resolution, with a single low-resolution camera and fixed Macro Block dimensions. Two are the implementations provided, with OpenMP and in CUDA on an NVIDIA Kepler GPU. Tests have been conducted on five image sequences and the results show a considerable improvement of the CUDA solution in all cases. Consequently, it can be stated that GPUs can efficiently accelerate computational times assuring the same image quality.