Optical eigenmode imaging (OEI) is a computational imaging method based on the expansion and projection of eigenmodes within an optical system. It acquires target images by characterizing and reconstructing spatial distribution information in the mode space. However, under low sampling rates, traditional OEI reconstruction is fundamentally limited by modal incompleteness, leading to the loss of high-frequency information. This physically manifests as enhanced side-lobes in the system’s point spread function and severe artifacts in the reconstructed images, thereby compromising edge sharpness and detail fidelity. To overcome these physical limitations, we propose a physics-driven, complex-valued convolutional neural network framework that seamlessly integrates physical laws with deep learning by incorporating the optical imaging physical model as an intrinsic constraint. Unlike conventional data-driven approaches that rely on massive paired datasets, our method utilizes an untrained complex U-net optimized through a self-supervised strategy. The network treats the real and imaginary components of the optical field as independent channels to strictly preserve phase information. Furthermore, to specifically address the issue of detail loss caused by modal incompleteness, we integrate a convolutional block attention module (CBAM) into the network’s decoder. This mechanism achieves high-quality image reconstruction through adaptive feature weight adjustment.
Numerical simulations indicate that under extremely low sampling rates, where traditional OEI fails to recover meaningful structural details and the baseline U-net suffers from residual artifacts, our attention-enhanced model successfully suppresses side-lobe-induced artifacts. Comparative validations demonstrate that superior image fidelity and structural similarity are achieved compared to traditional OEI algorithms and baseline deep learning approaches, effectively mitigating structural degradation caused by modal incompleteness. This method provides a robust solution for OEI, enabling high-quality super-resolution imaging without the need for high sampling redundancy or external training datasets.