45th Symposium on Computer Technology of Information, Systems and Applications
Conference site
Japanese:
English:
Tokyo
Abstract
Extracting and archiving the key characteristics of traditional landscapes is extremely important for various reasons, such as preservation and restoration. However, manual extraction of such characteristics has demerits, as it is both timeconsuming and subjective, to some extent. Automatic extraction of these, on the other hand, can potentially be both time efficient and consistent. Further, it may identify features overlooked by human experts. Motivated by this opportunity, we aim to enable such automatic extraction of characteristic features of cultural landscapes using artificial intelligence. In this study, we generated new cultural landscape images in an Island in Okinawa using a deep learning model (GAN) and then attempted to identify characteristic features of these using auto-generated activation maps (Grad-CAM). The activation maps reflect the regions that were used by the learning model in classifying an image as Izena-like, and are therefore informative regarding characteristic features, especially in generated images which closely resemble the original landscape. As a main result, we confirmed that the features of the streetscape can be extracted, and that the landscape features of streets that may have existed in the past can be made apparent by analyzing the generated images.