報告題目：Image Super Resolution with Generative Adversarial Networks
報告摘要：Image super resolution (SR) is a classical method to enhance the image visual quality. SR methods based on Generative Adversarial Networks (GAN) have the potential to recover realistic textures and missing details. In this talk, I will introduce four representative GAN based SR methods. As the seminal work, SRGAN first applies GAN for image SR in CVPR2017. Later in CVPR2018, SFT-Net incorporates image segmentation maps as additional priors to generate semantically meaningful details. After that, PIRM2018 held an SR Challenge and employed Perceptual Index (PI) to evaluate the perceptual quality. Enhanced SRGAN (ESRGAN) won the challenge on perceptual scores and became a new state-of-the-art. In ICCV2019, RankSRGAN further surpasses ESRGAN by employing a ranking network. It can be easily optimized in direction of indifferentiable perceptual metrics.
報告專家簡介：Chao Dong is currently an associate professor in Shenzhen Institute of Advanced Technology, Chinese Academy of Science. He received his Ph.D. degree from The Chinese University of Hong Kong, advised by Prof. Xiaoou Tang and Prof. Chen-Change Loy. In 2014, he first introduced deep learning method -- SRCNN into the super-resolution field. This seminal work was published in TPAMI and was chosen as one of the top ten “Most popular Articles” in 2016. His team has won the first place in international super-resolution challenges –NTIRE2018, PIRM2018 and NTIRE2019. He worked in SenseTime from 2016 to 2018, as the team leader of Image Quality Group. He, with his team, developed the first deep learning based “digital zoom” for smart phone cameras. His Google citation has surpassed 4900. His current research interest focuses on low-level vision problems, such as image/video super-resolution, denoising and enhancement.