증강 현실에서 선박 블록 모델 초기 정합을 위한 세그멘테이션 보정 연구

초록

In the process of extracting feature points of internal members within ship blocks for developing a block model alignment system with error detection, segmentation techniques were employed. However, the predicted masks generated through segmentation were often larger or smaller than the regions of interest, or contained noise in the form of empty spaces. Such noise can cause errors in the feature point?based initial alignment system for ship block models. In this study, we propose a methodology to refine segmentation model predictions by correcting these types of noise, and evaluate the performance of the correction method. Noise in the predicted masks was defined based on size and the presence of empty spaces, and an algorithm-based correction technique was designed to effectively remove and adjust such noise. The proposed correction approach was then applied to various ship block images, and the quality of the corrected masks was quantitatively assessed.