Enhancing dynamic target reconstruction and tracking based on ghost imaging and deep convolutional neural networks

Abstract

Ghost imaging requires a large amount of sampling data, which limits its applications in the study of dynamic objects. Here, we propose an imaging technique based on deep convolutional neural networks (SaDunet) that can be used to examine the dynamics of target objects. By replacing the traditional correlation imaging reconstruction approach with SaDunet, the ability to recover high-quality images at low sampling rates is enhanced. The motion process of the target object is decomposed into multiple motion frames, and then each frame is imaged separately. Experiments show that the reconstructed image of the target object obtained by this scheme is of high quality, contains almost no noise, and accurately reflects the motion behavior of the target object.