HoloPASWIN: Physics-Aware Swin Transformer

HoloPASWIN is a deep learning framework designed to eliminate twin-image artifacts in inline digital holography. By integrating the Angular Spectrum Method (ASM) with the Swin Transformer's global hierarchical attention, this U-Net architecture seamlessly and directly recovers physically-consistent amplitude and phase components from single intensity holograms.

Instructions:

• Upload your hologram (as a 2D intensity image).
• The network expects holograms calibrated at: Wavelength = 532nm, Pixel Pitch = 4.65µm, and recorded at Distance = 20mm.
• Due to the rigid nature of holographic physics parameters, the image will be center-cropped (not arbitrarily resized) to 224x224 prior to prediction.

Official Links:

• Hugging Face Model
• Official Code Repository
• Full Dataset