On global transform preservation by region based interest points for image registration

摘要

There are many methods to find Interest Points (IPs) in images for image registration. However, the underlying heuristics for finding them is different for each. Due to this, their behavior towards different image distortions is expected to vary. Through this paper, we attempt to investigate the truth of the following hypothesis - "Global Transform is better preserved by region based IPs as compared to Point based IPs". For this purpose, we make the use of "Speeded up Robust Features (SURF)" based IPs (which are point based) and "Maximally Stable Extremal Region (MSER)" based IPs (region based). Then by using "Random Sample Consensus (RANSAC)" on a standard stereo database (which is globally distorted afterwards), we validate the truth of the hypothesis we have made. The testing of this hypothesis is motivated from the fact that generally in medical images, global (usually affine) distortions are dominant. Local distortions tend to decrease registration accuracy if IPs are located at those sites. However, opposite is the case with temporally separated images (e.g. pictures of highway taken at an interval of 10 seconds, keeping camera fixed). They have dominant local distortion. Hence a proper choice of interest points for registering images is necessary.