多値の初期確信度を用いた建造物画像の自動GrowCut

抄録

In this paper, we propose an automated GrowCut method for building extraction from scenery images. GrowCut achieves a high-performance for image segmentation to a wide variety of images,but its quality of segmentation depends largely on accuracy of user-specified seed pixels. Therefore, if seed pixels are specified automatically and accurately, GrowCut can be employed as a promising method for automatic building extraction. The proposed method uses a priori knowledge that target is a building, in order to specify seed pixels automatically and accurately. Moreover, it also achieves a higher repeatability in building extraction than the conventional method by introducing multi-level seed strength in GrowCut.