Enhanced aberration correction in minimalist optical systems with the deep attention Wiener network

Aberrations in minimalist optical imaging systems present significant challenges for achieving high-quality imaging. Traditional methods often rely on precise aberration models, while deep learning approaches typically do not incorporate prior knowledge and lack interpretability. To address these limitations, we introduce the deep attention Wiener network (DAWNet), a differentiable framework that combines deep learning with Wiener deconvolution. DAWNet begins by extracting deep features using a convolutional neural network (CNN). These features are segmented by the field of view (FOV) based on spatial degradation and processed through a Wiener deconvolution module. To minimize distinct boundaries between adjacent segments, we employ innovative sliding sampling and a weighting matrix, facilitating efficient and accurate aberration correction closely aligned with physical models. The processed image blocks then pass through a post-processing attention module for further aberration correction and feature fusion, ultimately reconstructing high-quality images. We validate the efficacy of DAWNet by designing and implementing a minimalist optical system with two lenses. Both simulation and experimental results demonstrate that DAWNet significantly improves aberration correction compared to other techniques.