The use of deep learning for computational optical imaging: from data driven to physics driven

Recently deep neural networks (DNN) has shown the great capability of solving various inverse problems in computational optical imaging. Conventionally, DNN should be trained by a large set of paired or unpaired data. The most critical issue with this paradigm is that the neural network inference has no physical interpretation or limited generalization. In order to resolve these issues, one solution is to incorporate the physics of the problems in hand into the training of DNN, resulting in a novel framework that is called physics-enhanced deep neural networks or, PhysenNet, for short. Here we present a brief review of recent works in this regard with the use cases of phase imaging and ghost imaging. We will show that PhysenNet does not need any data to train. Instead, it learns from the physics of the given problem. Therefore the output of PhysenNet is naturally satisfied with the constrain imposed by the corresponding physical model. We would also like to emphasize that the concept of PhysenNet is quite generic, and can be used to solve many other inverse problems even outside the scope of imaging.