HoloPASWIN: Robust Inline Holographic Reconstruction via Physics-Aware Swin Transformers
A new AI model trained on 25,000 samples suppresses artifacts that have plagued lensless imaging for decades.
A team of researchers has introduced HoloPASWIN, a novel deep learning framework designed to solve a fundamental problem in inline digital holography (DIH). DIH is a valuable, lensless imaging technique prized for its simplicity and high-throughput capability, but it has long been plagued by the 'twin-image' artifact—a defocused conjugate wave that severely degrades reconstruction quality. While deep learning has been applied to this phase retrieval challenge, conventional Convolutional Neural Networks (CNNs) struggle with the global, long-range dependencies inherent in diffraction patterns. HoloPASWIN addresses this by leveraging a Swin Transformer architecture, which uses hierarchical shifted-window attention to efficiently capture both local details and these essential global relationships.
The technical innovation lies in its physics-aware design. The team developed a comprehensive loss function that integrates frequency-domain constraints with physical consistency, enforced via a differentiable angular spectrum propagator. This ensures the model's reconstructions are not just visually plausible but also spectrally accurate. The framework was rigorously validated on a large-scale synthetic dataset of 25,000 samples, incorporating diverse real-world noise types like speckle, shot, and read noise. The results demonstrate effective twin-image suppression and robust reconstruction quality, marking a significant step toward reliable, high-fidelity holographic imaging for applications in biomedical diagnostics, microscopy, and industrial inspection where artifact-free images are critical.
- Uses a Swin Transformer architecture to capture global diffraction patterns, overcoming the local receptive field limit of CNNs.
- Integrates a physics-based loss function with a differentiable angular spectrum propagator for spectrally accurate reconstructions.
- Trained and validated on a large-scale dataset of 25,000 synthetic samples with multiple noise configurations for robustness.
Why It Matters
Enables clearer, more reliable lensless imaging for medical diagnostics and lab-on-a-chip devices by solving a decades-old artifact problem.