Single-Shot Diffraction Autofocusing: Distance Prediction via an Untrained Physics-Enhanced Network

Abstract

Deep learning technology has shown excellent performances and successful applications in optical information processing. However, the long-time training, large amount of manually labeled data and generalization capability hinder the application of deep neural network (DNN) under supervised learning. The deep image prior (DIP) opinion promotes the development of untrained neural network, which can learn from one image. Here we propose a DIP-based strategy to nest the DNN into a physical model for finding the optimal solution in a univariate optimization problem. The untrained physics-enhanced network (UPN) is proposed to predict the diffraction distance via only one diffraction pattern of a known phase object. Simulation and experimental results show that the UPN can be used to predict the distance precisely and consistently with different targets, diffraction distances as well as phase ranges, while it only takes a little time for training. In addition, the trained UPN can generalize to the other targets as long as the actual diffraction process keeps the same. Compared with the autofocusing metrics of holographic reconstruction and traversal method, the UPN has advantages in speed and accuracy, and it also has good noise resistance, which are all meaningful for the autofocusing of holographic reconstruction and imaging.

Keywords

• Deep image prior
• prediction of diffraction distance
• physical model
• untrained physics-enhanced network