FPM-DCN: Fourier ptychographic microscopy with a dual-constrained network for robust reconstruction in complex...

As a computational imaging technique with significant application prospects, Fourier ptychographic microscopy (FPM) has achieved considerable success through integration with deep learning, yet its dependence on large training datasets limits generalization, and it is challenging to achieve robust reconstruction performance in complex imaging conditions with significant noise and aberrations. This paper proposes FPM with a dual-constrained network (FPM-DCN), a weakly-supervised network integrating dual constraints through an embedded Fourier ptychographic physical model and a non-dense matching perceptual loss. A physics-constrained reconstruction network enhances generalization with few-shot training, while the perceptual loss fine-tunes parameters by learning from brightfield modality information to reduce noise and improve contrast. An aberration correction network with spatial attention recovers Zernike coefficients and refines predictions under the physical model constraint. Leveraging a hybrid data-model-driven strategy, FPM-DCN achieves robust, high-contrast, and low-artifact reconstruction under diverse complex imaging conditions.