Iterative random search: a new local minima resistant algorithm for motion estimation in high-definition videos

Abstract

Motion estimation (ME) is the most important component of current video encoders, however, it presents a very high computational complexity. To deal with this complexity, fast ME search algorithms are widely used, since they can greatly speed up this process. Fast search algorithms are vulnerable to choose local minima, producing quality losses, and these losses are more significant when high-definition videos are considered. This work presents a new fast search algorithm for motion estimation, focusing on high-definition videos, named Iterative Random Search (IRS). The IRS algorithm randomly chooses candidate blocks from the reference frame, and, for the best candidates, an iterative refinement is done. The central position is also evaluated through an iterative process. By using this combination of strategies, the IRS becomes less susceptible to local minima falls. Achieved results show that, for 1080p sequences, IRS generates the highest quality results when compared to well known fast algorithms, such as Diamond Search, Four Step Search and Three Step Search. The quality gains can be higher than 4 dB, while the number of evaluated candidate blocks may increase, at most, 2.6 times. Additionally, the average quality loss against Full Search is 1.45 dB, while the number of evaluated candidate blocks can achieve a reduction higher than 200 times.