Deep learning dynamically orders Hadamard patterns for sharper spectral images

Single-pixel imaging (SPI) offers a low-cost alternative to expensive spectral sensors, but its quality has been limited by pre-designed, fixed optical coding patterns. Current Deep Optical Coding Design (DOCD) methods train a fixed pattern offline, which only works well for scenes similar to those in training data. This paper introduces a scene-driven ordering of Hadamard basis patterns, where the selection is optimized end-to-end based on the actual scene characteristics during acquisition. The method reformulates the DOCD framework to allow flexible, adaptive pattern ordering, using the fact that SPI typically acquires hundreds of snapshots.

By learning to order the Hadamard matrix dynamically for each scene, the approach significantly improves the reconstruction quality of both visible (VIS) and near-infrared (NIR) spectral images compared to fixed designs. The authors validated their method with simulations on spectral datasets and real test-bed acquisitions. This advancement could make high-resolution spectral imaging more accessible for precision agriculture, environmental monitoring, and other applications where traditional sensors are too costly.