End-to-end optimization of illumination patterns for ultra-low sampling single-pixel target recognition

Advanced imaging methods such as Single-pixel imaging (SPI) are subjected to a fundamental trade-off between the sampling sparsity and the fidelity of target recognition, which significantly constrains their applicability in photon-starved environments including low-light fluorescence microscopy and imaging through scattering media. To address this issue, we present an end-to-end optimized SPI apparatus that integrates data-driven deep learning (DL) with physical constraints. Our key innovation is a two-level optimization-based approach where we jointly optimize the ultra-sparse speckle patterns and a DL-based classifier using Convolutional neural network (CNN) for synthetic training samples, while physics-driven refinements tailor these patterns to real-world experimental constraints. With 99% accuracy at 0.10 sampling and 92% accuracy at an extreme 0.06 sampling rate, four times sparser than traditional methods, the optimized patterns exhibit exceptional hardware compatibility and a superior sparsity-performance trade-off. A unified simulation-to-experiment training paradigm that successfully closes the reality gap is one of the major developments. With direct applications to biological imaging and non-destructive material characterization, this study lays the groundwork for photon-efficient computational imaging systems where speed, photon efficiency, and accuracy are crucial.