An analysis of sub-pixel estimation error on area-based image matching

Abstract

Area-based matching and sub-pixel displacement estimation using similarity measures are common methods for various fields. Sub-pixel estimation using parabola fitting over three points with their similarity measures is also a common method to increase the resolution of matching. However, this estimation contains a systematic error depending on the image characteristics, the similarity function, and the fitting function used for the sub-pixel estimation. The characteristics of sub-pixel estimation error have been clarified using a simple analysis model. Although this model is quite simple, it enables us to make a theoretical analysis for general images without assuming any specific image pattern. Through this analysis, it has been shown that there are the right combinations of similarity functions and fitting functions. Also, the so-called "pixel-locking" effect, where the estimated positions tend to be biased toward integer values, is explained. Finally, the analyses are verified with an experiment using real images.