Great question! This topic is still under investigation by a number of research groups. The short answer is, around 60-70% overlap works very well, if you are ok with such measurement redundancy. 50% overlap is near the limit to get acceptable results for general samples, and then it is possible to use even less overlap if one applies prior knowledge about the sample (i.e., constraints their search to a subset of possible samples)

Yes we have worked on that a bit in the past. In the paper below, we integrated an untrained convolutional neural network as a "deep image prior" to help improve our FP reconstruction results (for the 3D imaging case). In the second link below, we used a machine learning framework to optimize the LED pattern within our FP microscope. A GAN approach sounds interesting! opg.optica.org/oe/fulltext.cfm?uri=oe-28-9-12872&id=430210 opg.optica.org/boe/fulltext.cfm?uri=boe-10-12-6351&id=423327

Hi, I made this video, but it is based on number of different published papers. Below are some useful introduction and review papers that we have recently written about Fourier ptychography: K. C. Zhou et al, "Introduction to Fourier Ptychography: Part 1," Microscopy Today (2022) Link: bit.ly/3LdKQ7P P. C. Konda et al., "Fourier ptychography: current applications and future promises," Opt. Express 28, 9603-9630 (2020) Link: opg.optica.org/oe/fulltext.cfm?uri=oe-28-7-9603&id=429273

@@dukemicroscopes Very nice video. Thankyou so much