Physics-Guided Multi-Frame Image Reconstruction: From Simple Deep Models to Phase Retrieval Integration

Abstract

Optical aberrations present a fundamental challenge in optical imaging systems, significantly degrading image resolution and quality by corrupting Fourier phase information. While hardware-based solutions such as Adaptive Optics (AO) can compensate for these distortions, they require complex and expensive instrumentation. Alternatively, computational approaches such as Coherent Averaging (CA) exploit multiple distorted images to recover phase information without additional hardware. However, CA typically requires a large number of input frames and exhibits limited robustness in the presence of complex or spatially varying aberrations. In this thesis, we propose a physics-guided multi-frame deep learning framework for optical image reconstruction that integrates the physical principles of CA with modern deep learning architectures. Rather than relying on a single distorted observation, the proposed approach leverages frame-to-frame diversity across multiple aberrated images to enable data-efficient Fourier phase recovery. We first conduct a controlled study using a simple U-Net architecture to verify the efficacy of multi-frame inputs, demonstrating that reconstruction performance improvements arise primarily from increased input diversity rather than model complexity. Through this background, we introduce a transformer-based Shift-correction network that learns inter- frame shifts and incorporates them into an iterative reconstruction workflow inspired by CA. To systematically evaluate the proposed method, we construct a staged synthetic dataset based on Zernike polynomial–based aberrations, ranging from simple global shifts to complex global aberrations and spatially varying local distortions. Experimental results show that the proposed network achieves diffraction-limited reconstruction performance comparable to conventional CA in scenarios involving global shifts and aberrations. In the results reveal inherent limitations when addressing local aberrations, indicating that global shift correction alone is insufficient in such scenarios. These findings motivate future extensions toward multi-stage frameworks that decouple local aberration correction from global alignment. Overall, this work demonstrates that physics-guided multi-frame deep learning provides a viable and data-efficient computational pathway for diffraction-limited image reconstruction under local aberration imaging conditions, serving as a complementary alternative to hardware-based aberration correction.