4× sparser
End-to-end optimized illumination patterns achieve 92% recognition accuracy at an extreme 0.06 sampling rate with exceptional hardware compatibility, outperforming traditional methods in the sparsity-performance trade-off.
Ayesha, Abbas · Cao, Jie · Saba, Sajid · Muhammad, Qasim · Tan, Wang
Optics and Laser Technology 2026
Sampling rate as low as 0.06 still yields 92% accuracy, requiring four times fewer samples than conventional approaches.
Optimized patterns are directly compatible with existing hardware, refined with physical constraints to load onto a spatial light modulator without extra calibration.
A unified simulation-to-experiment training paradigm successfully closes the reality gap, enabling direct transfer of simulation-trained patterns to experiments.