End-to-end learned optics for IR under-display lensless face recognition via digital-twin simulation

We present an infrared-assisted under-display lensless face recognition system evaluated using a simulation-based digital twin, in which a coded mask and a deep neural network classifier are jointly optimized end-to-end. The sensor-level mask converts facial images into visually unreadable measurements, preserving privacy at capture. We compare heuristic and learned mask designs under a fixed transmittance constraint across multiple datasets and show via MTF analysis that the learned mask adapts its frequency response during training, leading to superior recognition. The optimized mask achieves 88.67% accuracy on DigiFace1M, slightly outperforming all heuristic baselines and highlighting the promise of data-driven optical co-design in computational imaging.