Deep Learning Reconstruction for Single Pixel Imaging with Generative Adversarial Networks

Single pixel imaging (SPI) enables high-resolution imaging through multiple coded measurements based on low-resolution snapshots. An inverse problem can then be solved to reconstruct a high-resolution image given the coded measurements. There has been recent interest in adoption of deep neural networks in SPI reconstruction. However, existing methods are commonly trained with pixel-wise loss terms such as the ℓ 1 -norm loss, which can result in spatial blurring and poor sensitivity to structural details. In this study, we propose a novel approach for deep SPI reconstruction based on an unrolled conditional generative adversarial network (cGAN) model. The generator estimates the high-resolution image using coded low-resolution measurements by iterating across a cascade of denoising and data-consistency modules. Meanwhile, the discriminator distinguishes real versus synthesized high-resolution images. The architecture is trained end-to-end via a combined pixel-wise and adversarial loss to enhance sensitivity to structural details. The proposed method is demonstrated against existing SPI reconstruction methods, and ablation studies are performed to demonstrate the individual model components. The proposed method outperforms competing methods in terms of both quantitative metrics and visual quality.