Fast adaptive parallel computational ghost imaging based on meta learning

Computational ghost imaging has emerged as a powerful technique that generates images by interrogating objects with a series of illumination patterns. However, multiple measurements in the temporal domain are required to obtain high-quality images. Data-driven one-shot learning has been proposed to reconstruct satisfactory images from undersampled measurement signals, but their inflexibility hinders practical application. In this paper, we propose a meta-learning-based parallel computational ghost imaging to overcome the trade-off between data acquisition time, adaptation time, and image quality. Compared with the general sampling, the sampling speed can be increased several times by multiplexing the time-varying patterns. Additionally, a two-stage learning strategy reduces the time and cost of retraining the model from scratch when the system configuration changes (e.g., illumination patterns, sampling rate). Consequently, the proposed method significantly enhances the practicality of computational ghost imaging, providing an optional solution for real-time imaging.