Shuhe Zhang：Understanding data distribution for better computational imaging

Embracing your dream ideal, you are scratched by what is real! — a perfect optical system simply doesn't exist in this world. But fear not! Computational imaging, our digital "magician," is breaking the physical limits of traditional optics with its "hardware + algorithm" combo. Thanks to advances in machine learning and optimization theory, we can now frame computational imaging tasks through Bayesian models. Yet, real-world imaging physics is often oversimplified, leading to performance bottlenecks—mostly due to inadequate modeling of the joint light field-data distribution. From early iterative phase retrieval to today’s end-to-end deep learning methods, most approaches either oversimplify optical degradation as a linear process or rely on data-driven "black-box" training, lacking a physical understanding of the nonlinear coupling between light propagation and sensor response. The result? Reconstructed images plagued by artifacts, resolution loss, or noise amplification—especially in challenging conditions like low light, wide fields of view, or system uncertainties. In this talk, I’ll unveil our group’s breakthroughs in holographic imaging and their computational applications: (1) Decoding the "Secret Sauce" of Computational Imaging – Starting from maximum a posteriori (MAP) estimation, we derive the "three pillars" of computational imaging: forward modeling, inverse problem design, and optimization. We theoretically explain how noise distribution modeling shapes the imaging process. (2) Feature-domain Phase Retrieval – Instead of obsessing over every pixel, our architecture leverages multi-scale image features to guide phase recovery, overcoming forward model limitations even with significant system errors. (3) Latent-wavefront Phase Retrieval – Using an EM-style "alternative optimization" approach, we alternately tackle non-convex phase retrieval and convex subproblems, enabling fast, high-quality imaging. This paves the way for label-free cell imaging and wide-field pathology analysis. Dr. Shuhe Zhang is currently a Postdoctoral Research Associate in the Department of Precision Instrument at Tsinghua University, where he was awarded the prestigious Tsinghua "Shuimu Scholar" Postdoctoral Fellowship. He obtained his Ph.D. in 2023 through the Chinese Scholarship Council, having conducted his doctoral research in the Netherlands since 2019. Dr. Zhang specializes in holographic optical imaging and medical image processing, with a particular focus on advancing the integration of holographic optics with medical engineering applications. His research has been published in journals including Advanced Science, Optica, Laser & Photonics Reviews, and Medical Image Analysis.