Examination and Issues of Kumamoto Castle Ishigaki Region Extraction Focusing on Stone Contour Features

Abstract

An automated method for extracting stone contours from stone-wall images is examined in the context of the reconstruction of Kumamoto Castle. Various methods are considered: the GrabCut method, which uses a background label for highly-linear parts considered to be features of stone contours, a variant of the GrabCut method that uses low-brightness regions instead of high-linearity parts as background labels, a watershed-based method, and a method combining the GrabCut and watershed methods. In this paper, we report the extraction results obtained by each method. For each stone-wall image, the method yielding the smallest error relative to the ground truth is applied to provide a contour candidate. The results show that the line-segment GrabCut, which considers high linearity, is effective for many stone-wall images. Other methods are also effective, depending on the stone-wall image considered. The best results were obtained with an accuracy of approximately 40% to 50% within the upper limit of the allowable error. Finally, we consider outstanding issues. Owing to the low accuracy of the contour feature extraction, there is contamination of the target region with some of the surrounding region. Another problem is that it is divided extensively by line segments extracted in the texture of the stone surface. In the future, we will seek to improve contour extraction using deep learning and develop an index for matching stone shapes.