Minimalist optical system image restoration based on deep attention Wiener network

Aberrations in minimalist optical imaging systems pose significant challenges to achieving high-quality imaging. Traditional Wiener filtering methods, though effective, are constrained by their dependency on precise blur kernels and noise models, and their performance degrades with spatial variations in these parameters. On the other hand, deep learning techniques often fail to fully utilize prior information about aberrations and suffer from limited interpretability. To address these limitations, we propose a novel deep attention Wiener network (DAWN). This approach integrates deep learning with Wiener filtering to enhance image restoration while reducing computational complexity. By using optical simulations to generate blur kernels and noise models that closely mirror real conditions, our method fits distinct point spread function (PSF) for different fields of view (FOV), creating a robust dataset for training. The DAWN model first employs a convolutional neural network (CNN) for feature extraction, followed by sequential Wiener filtering applied in half FOV block length steps. To further improve image restoration, a nonlinear activation free net (NAFNet) is used to correct discrepancies introduced by simulated blur kernels and noise models. The model is trained end-to-end, and to streamline the process, Wiener filtering is confined to 4 × 4 FOV blocks. A weighting matrix within the Wiener filtering layer mitigates seams between adjacent blocks. Simulation and experiment results demonstrate that our approach outperforms the mainstream image restoration methods.