A STUDY OF ANNOTATION METHODS FOR A SMALL NUMBER OF PHENOMENON USING DEEP LEARNING

Abstract

In this paper, we study annotation methods to improve the accuracy of semantic segmentation for a small number of phenomenons in civil engineering. The target of the segmentation is dead trees. The annotation methods to be considered are those that annotate only dead trees and those that annotate not only dead trees but also artificial structures. As a result of comparing these two annotation methods, we found that the extraction accuracy of the latter method was higher than that of the former, and that the extraction accuracy was improved by annotating objects that are easily misread. From the point of view of dead tree interpre-tation, it was also found that this improvement in extraction accuracy saves labor and awareness for the engineers.