Fundamental Research about Identifying Accurate Corresponding Points using Image Super-Resolution

Abstract

Corresponding points search is an indispensable process of stereo measurement. In this research, we suggest a method to reduce corresponding point error using deep learning. We found that upscaling was effective for binary markers the size of which was from 20 to 180 pixels. Although the amount of error reduction changed depending on the shape, rotation angle and size of markers, the results showed that the error was always reduced. The effectiveness was confirmed in both computer simulation and practical measurement.