Dual-attention cGAN for high-fidelity ghost imaging at ultra-low sampling rates

Achieving high-quality ghost imaging (GI) reconstruction under extremely low sampling rates remains a significant challenge. This paper proposes a novel deep learning-based approach for high-fidelity GI reconstruction. Leveraging an end-to-end conditional generative adversarial network (cGAN), the method incorporates residual structures integrated with efficient channel attention (ECA) and convolutional block attention module (CBAM) mechanisms to significantly enhance feature extraction and representation. This framework enables direct reconstruction of target images from one-dimensional bucket detection signals. Pre-trained on simulated datasets, the proposed method demonstrates exceptional performance. The results under ideal simulation conditions demonstrate that the proposed method can achieve high quality (SSIM \geq 0.8) even at the extremely low ratio of 0.29%. Experimental results on the MNIST and Fashion-MNIST benchmark datasets confirm that accurate image reconstruction can be obtained at remarkably low sampling rates (0.78%). Compared to traditional computational GI, baseline cGAN methods and GAN with integrated SENet attention module consistently demonstrate the superiority of our approach, establishing a new state-of-the-art in low-sampling-rate GI reconstruction.