2021 Volume 2 Issue J2 Pages 510-516
A number of studies on generative adversarial networks (GAN) has been conducted to improve work efficiency in various fields. Focusing on the land readjustment project, it may be possible to utilize GAN in image transformation such as pix2pix. The objectives of this study are to confirm the learning status of pix2pix regarding the basic design conditions in the design of land replotting, and to verify the possibility of using pix2pix to improve the work efficiency of the design of land replotting. This study prepares datasets with a unified area scale based on the land shape at right angles and parallels. Numerical experiments are performed to investigate transformed images of validation data by pix2pix after training. Results show images generated from pix2pix-based transformation contain information on replotting of land polygons with various size and shape. Thus, it suggests that pix2pix learns explicit knowledge on land replotting in training process.
