HSR-CFAT-based small-sample super-resolution reconstruction framework for off-axis digital holography.

Area of Science:

• Optics and Photonics
• Computer Vision
• Artificial Intelligence

Background:

• Digital holography captures 3D object wavefronts, but sensor limitations (pixel size, diffraction) yield low-resolution holograms.
• Deep learning super-resolution typically requires large datasets, which are difficult to acquire due to hardware and experimental constraints.

Purpose of the Study:

• To develop a novel small-sample super-resolution reconstruction framework for digital holography.
• To overcome the dependency on large datasets for hologram super-resolution.

Main Methods:

• Proposed the holographic super-resolution CFAT (HSR-CFAT) model, incorporating a residual hybrid attention group (RHAG).
• RHAG fuses hybrid attention blocks (HAB_D, HAB_S) and overlapping cross-attention (OCAB) with sliding window attention (SWA) for multi-scale feature extraction.
• Integrated a phase-aware upsampling (PAU) module for computational efficiency and phase accuracy, enabling pixel-level reconstruction.

Main Results:

• A small dataset of 377 multi-view holograms was used for training and validation.
• The HSR-CFAT framework successfully performed super-resolution processing directly on holograms, generating high-resolution phase images.
• Achieved significantly improved subjective visual quality and objective evaluation metrics compared to existing methods.

Conclusions:

• The HSR-CFAT framework offers an effective solution for super-resolution reconstruction in digital holography using small datasets.
• This work presents a new technical approach for off-axis digital holographic imaging, overcoming sensor pixel limitations.
• Demonstrates robust performance and overcomes the reliance on large datasets for high-quality hologram reconstruction.