Image processing for shape transformation using CycleGAN
Akira Nakajima, Hiroyuki Kobayashi · The Proceedings of JSME annual Conference on Robotics and Mechatronics (Robomec) · 2023
CycleGAN is a technique that realizes image transformation by learning the relationship between the domains of two images. CycleGAN is good at style conversion such as color and pattern, but there is a problem that conversion accompanied by shape conversion is difficult. The reason is that CycleGAN recognizes the image background as part of the conversion target and cannot perform feature extraction well. In this research, shape transformation by CycleGAN is performed by preparing a data set in which the image background is deleted by a generative model.