题 目：Image recovery: interaction between regularization and learning
摘 要：Image recovery is about restoring clear images from degraded ones. It usually requires solving an challenging ill-posed inverse problem. Regularization has been one main tool for image recovery by imposing certain prior on clear images via certain variational form. In this talk, we will discuss how regularization methods and machine learning techniques can interact each other to have the start-of-the-art performance in image recovery. For the illustration, two applications are discussed in this talk. The first is how Bayesian learning can be introduced in the existing regualarization methods for blind motion deblurring of natural images. The other is how-norm relating regularization can lead a collaborative deep learning method for super-resolving degraded text images. Benefiting from such interactions, both approches outperformed existing ones by a large margin.