Beijing Institute of Mathematical Sciences and Applications

Yunfeng Cai

Yunfeng Cai

Thursday, April 24, 2025 2:00 PM - 3:00 PM

A3-4-312

In this talk, a new semantic-guided coarse-to-fine generative image inpainting network (SCF-GMIN) is proposed to solve this challenging problem. The main idea is to construct a new two-stage conditional generative adversarial network to fill the missing regions involving multiple semantic categories by introducing external semantic information. Instead of directly inpainting the corrupted image, we leverage the semantic map (SM) inpainted in stage one to guide the inpainting process of corrupted regions in stage two. Both two stages share a coarse-to-fine network that enhances the visual quality of restorations. In our model, the proposed feature module plays a crucial role by storing semantic information of different categories in the feature space with the help of SM. The image inpainting tasks of semantic maps and corrupted images complement each other in the training process, which promotes a balance between understanding image semantics and maintaining visual consistency.

Yunfeng Cai studied in Mathematics at the University of Science and Technology of China from 2000 to 2004. He then pursued his PhD in Computing Mathematics at Peking University, which he obtained in Jan. 2009. From Jan. 2009 to June 2012, he worked as a postdoctoral researcher at the Academy of Mathematics and Systems Science / University of California, Davis. From Sep. 2012 to Sep. 2018, he served as a researcher at Peking University. In Sep. 2018, he joined Baidu Research as a Research Scientist and left on May 2024. In June 2024, he began his current role as a Professor at the Beijing Institute of Mathematical Sciences and Applications (BIMSA).