报告题目：Self-supervised Learning and Model Adaption for Solving Inverse Imaging Problems

报告时间：2024.12.10   10:00-11:00

报  告 人 ：Prof.  Ji Hui （National University of Singapore）

报告地点：雷军科技楼八楼报告厅（806）

Abstract：Deep learning has revolutionized many fields, including inverse imaging problems, but most solutions rely on supervised learning, requiring extensive ground-truth data for network training. This limits their use in data-scarce areas like medicine and science. In this talk, we present a series works on self-supervised learning for image reconstruction, where a DNN can learn to predicts images from noisy or incomplete measurements without seeing any ground-truth data. It is achieved by understanding DNN-based inversion from the perspective of Bayesian inference. Moreover, we also show how to efficiently adapt pre-trained models from one image domain to new image domains, fully leveraging the advantages of both supervised and unsupervised learning. Experiments showed that our method competes well against supervised approaches in many real-world tasks, and model adaptation can further boost performance to unprecedented levels.