Deep Learning Piston Sensing for Sparse Aperture Systems With Simulated Training Data

Abstract

The image-based piston sensing method using the convolutional neural network (CNN) is an advanced technique which has good applicability. However, acquiring a large amount of the training dataset required to train a network is difficult to handle in practice. In this letter, we demonstrate the possibility of using a neural network trained by the simulation dataset to accurately sense pistons directly from experimental images. As a demonstration of the proposed scheme, a single CNN developed by computer-generated images is applied for piston measurement of an experimental setup with three sub-apertures. This is particularly helpful for the sparse aperture system with more sub-apertures. We believe that the study in this letter will contribute to the applications of the CNN-based technique for piston sensing.