DynGhost: Temporally-Modelled Transformer for Dynamic Ghost Imaging with Quantum Detectors
Researchers Vittorio Palladino and Ahmet Enis Cetin have introduced DynGhost, a novel transformer architecture designed to advance dynamic ghost imaging using quantum detectors. Traditional deep learning methods for ghost imaging often fail to address two critical issues: the inability to exploit temporal coherence across video frames and the reliance on inaccurate additive Gaussian noise models that do not reflect the Poissonian statistics of real single-photon hardware. DynGhost resolves these limitations by employing alternating spatial and temporal attention blocks to capture dynamic scene changes effectively. Furthermore, the team developed a quantum-aware training framework utilizing physically accurate simulations of detectors such as SNSPDs, SPADs, and SiPMs, combined with Anscombe variance-stabilizing normalization. This approach corrects distribution shifts that typically cause classical models to fail under realistic hardware constraints. Experimental results across multiple benchmarks indicate that DynGhost significantly outperforms both traditional reconstruction techniques and existing deep learning architectures. The model demonstrates particular improvements in dynamic scenarios and photon-starved environments, marking a significant step forward in computational imaging and quantum sensing technologies.
Wire timeline
DynGhost: Temporally-Modelled Transformer for Dynamic Ghost Imaging with Quantum Detectors
Researchers Vittorio Palladino and Ahmet Enis Cetin have introduced DynGhost, a novel transformer architecture designed to advance dynamic ghost imaging using quantum detectors. Traditional deep learning methods for ghost imaging often fail to address two critical issues: the inability to exploit temporal coherence across video frames and the reliance on inaccurate additive Gaussian noise models that do not reflect the Poissonian statistics of real single-photon hardware. DynGhost resolves these limitations by employing alternating spatial and temporal attention blocks to capture dynamic scene changes effectively. Furthermore, the team developed a quantum-aware training framework utilizing physically accurate simulations of detectors such as SNSPDs, SPADs, and SiPMs, combined with Anscombe variance-stabilizing normalization. This approach corrects distribution shifts that typically cause classical models to fail under realistic hardware constraints. Experimental results across multiple benchmarks indicate that DynGhost significantly outperforms both traditional reconstruction techniques and existing deep learning architectures. The model demonstrates particular improvements in dynamic scenarios and photon-starved environments, marking a significant step forward in computational imaging and quantum sensing technologies.
cs.AI updates on arXiv.org