Fast High‐Fidelity Fourier Ptychographic Microscopy via Wavelet Transform and Linear Attention

ABSTRACT Fourier ptychographic microscopy (FPM) is an emerging computational imaging technique that breaks through the optical diffraction limitations, enabling large‐field, high‐resolution observation of biological samples. However, traditional FPM iterative reconstruction algorithms suffer from high computational complexity and poor real‐time performance. Additionally, existing data‐driven FPM models exhibit limited generalization due to simulation‐dominated training data. To overcome these challenges, the FPM‐BioCell dataset of multi‐source biological samples is constructed based on the previously proposed forward wavelet‐transform model family (WL‐FPM). Furthermore, a generative adversarial network integrating wavelet transform blocks and mamba‐inspired linear attention models, termed WM‐FPM, is proposed for FPM reconstruction. Through multi‐scale feature extraction, long sequence modeling, and adversarial training, WM‐FPM breaks the speed‐accuracy trade‐off in FPM reconstruction. It outperforms state‐of‐the‐art deep learning models on the FPM‐BioCell dataset with SSIM (0.71), MS‐SSIM (0.87), PSNR (28.55), FID (35.35), and LPIPS (0.27). Crucially, when validated against an independent optical ground truth, WM‐FPM not only generalizes effectively to unseen samples but also achieves a speedup of over 100× compared to the physics‐based WL‐FPM, reconstructing a full‐field 12288 × 12288‐pixel image in merely 5.63 s from a single 2048 × 2048‐pixel input, without compromising fidelity. The code and dataset are open‐sourced to facilitate the rapid deployment of FPM in biomedicine.