Robust optical flow estimation for underwater image

Abstract

This paper describes a novel and robust method of estimating optical flow for underwater image sequences. This method can give out reliable optical flows output, even if the quality of image is very poor. In order to estimate optical flow of current templtate, first, we divide the template into some sub-templates, and then copute the similarity profiles of each sub-template. These similarity profiles can be used to extract two voting roles: positive voting role and negative voting role. The positive voting role can be used to increase correct optical flows, and the negative voting role can be used to reduce incorrect optical flows. We use two voting qualification variable (TP, TQ) to control the positive and negative voting processing, and use a SNR value to evaluate the each voting result. The estimated optical flow is reliability when SNR value converge to infinity only by useing small TP and TQ.