Phase unwrapping based on self-attention dense residual network

Abstract

Phase unwrapping, a critical step in obtaining holographic information, plays a significant role in the field of digital holography, particularly in applications such as fringe projection for 3D imaging, synthetic aperture radar, and magnetic resonance imaging. Traditional phase unwrapping algorithms often suffer from error accumulation, high computational costs, and poor performance in low signal-to-noise ratio (SNR) environments. To address these issues, this paper proposes a novel deep learning framework, named as Self-Attention Dense Residual Network (SA-DRNet), for phase unwrapping. To obtain continuous phase, we initially employed a dense network for multiple extractions of phase features. However, to alleviate the phase discontinuities and phase jumps caused by gradient issues, we integrated residual connections within the dense network. Finally, we incorporated a self-attention module to enhance the global phase information restoration, including the background phase, thereby achieving high-precision phase acquisition. Additionally, we established an off-axis digital holographic optical system to capture the holograms of the USAF resolution test target and artificial ink dots. Finally, the robustness of the proposed algorithm under severe noise conditions was first verified through numerical simulations, followed by experimental validation of its effectiveness.