题    目：Exploring Diffusion-based Image Enhancement

主讲人：岳宗胜

单    位：新加坡南洋理工大学

时    间：2024年9月22日 9:00

腾讯ID：922-587-332

摘    要：Diffusion models have achieved unprecedent success in the field of image generation recently. Image enhancement is a more challenging problem beyond image generation due to its highly ill-posedness. In this talk, we delve into the innovative application of diffusion models for image enhancement and present our recent advancements along this research trend, mainly focusing on three core aspects including diffusion prior, diffusion model design, and diffusion loss. First, we discuss how to leverage diffusion prior to capture rich and realistic image details, leading to more perceptually appealing enhancements. Next, we examine the design of diffusion models tailored for image enhancement that improve the model efficiency and effectiveness. Finally, we introduce a novel diffusion loss to increase the models’ generalized capabilities across various domains. Through this exploration, we aim to highlight the potential of diffusion-based approaches in pushing the boundaries of image enhancement technologies.

简    介：新加坡南洋理工大学MMLab@NTU博士后。2017年至2021年在西安交通大学数学与统计学院孟德宇教授指导下攻读博士学位。2017年和2019年分别在香港中文大学（导师Leung Yee）和香港理工大学（导师Lei Zhang）担任科研助理。2021年至今分别在香港大学（合作导师Kenneth K.Y. Wong）和新加坡南洋理工大学(合作导师Chen Change Loy)从事博士后研究。主要研究方向为机器学习概率建模及其在底层视觉任务中的应用，包括但不限于图像去噪，图像超分辨和图像去模糊等。近五年来在人工智能国际顶级期刊和会议上发表多篇论文，包括TPAMI, IJCV, NeurIPS, CVPR等，Google Scholar引用1200余次。