Computational Ghost Imaging via Speckle Illumination and Deep Learning Under Low Sampling Rates

In recent years, ghost imaging (GI), an unconventional imaging technique, has garnered increasing attention from researchers. Its most notable feature is the ability to acquire information about an object using only a single-pixel detector, although the image quality heavily depends on the sampling rate. In this work, we propose a deep learning-based method for high-quality image reconstruction at low sampling rates. Specifically, we propose a novel speckle illumination pattern, which exhibit superior imaging capabilities at extremely low sampling rates, thereby enhancing image reconstruction quality. The deep learning approach employs an end-to-end conditional generative adversarial network (SE-CGAN), integrated with the SENet attention module. It reconstructs the target image solely from one-dimensional light intensity sequences, using simulated data for pre-training to achieve high-quality object image reconstruction. Results indicate that our method outperforms traditional computational ghost imaging and compressed sensing (CSGI) approaches, achieving high-quality image reconstruction even at very low sampling rates (as low as 0.2%).