主催: 公益社団法人精密工学会
会議名: 2020年度精密工学会春季大会
開催地: 東京農工大学
開催日: 2020/03/17 - 2020/03/19
p. 200-201
Material Segmentation for industrial 3D CT scans are critical for obtaining more coherent visualizations of a scanned object.Although manual and semi-automated segmentation methods exist, they tend to be time-consuming and require expertise. The theme of this research is to explore the possibility of implementing Convolutional Neural Networks (CNNs) to automate the segmentation process of a prespecified set of materials. However, one common obstacle associated with CNNs is the amount of training data it requires to converge; in addition, 3D CT data is scarce and difficult to obtain. To overcome this, this paper will adopt U-Net which is a convolutional network designed for biomedical image segmentation. U-Net presents a novel method to obtain 2D segmentations with very few training images. New architectures based on U-Net will be developed to exploit the spectral information from dual-energy CT to improve material segmentation accuracy.