Enhanced ghost imaging reconstruction via a Chambolle-Pock inspired deep unfolding network

Abstract

Ghost imaging is an innovative imaging modality that addresses the ill-posed reconstruction challenges associated with the acquisition of sparse measurements using a bucket detector. This technique holds extensive potential for applications and possesses significant practical utility across various domains. Deep unfolding networks (DUNs) based on compressive sensing techniques and deep learning methods have been applied to ghost imaging reconstruction in recent years due to their adaptive solid learning capabilities and inherent interpretability. However, most DUNs exhibit a lack of sensitivity to the reconstruction of image details at low sampling rates, resulting in the introduction of distortion and blurring in complex images. In this paper, we propose a ghost imaging method based on a deep unfolding network inspired by the Chambolle-Pock (CP) Algorithm. This method combines the CP algorithm with DUNs, enhancing the visual quality of reconstructed images while reducing the number of model parameters. Furthermore, we propose a multi-scale information mapping module for extracting and integrating the sensitivity of different scale feature information, thereby mitigating information loss in the reconstruction stage and improving image reconstruction details. The proposed method is shown to enhance the reconstruction quality of images, particularly in terms of detail recovery, and to outperform existing techniques.