Computational compressive imaging techniques such as single-pixel and lensless cameras reconstruct objects from far fewer measurements than traditional sensors require. However, the inverse reconstruction step usually relies on compressed-sensing algorithms or deep neural networks, and the electronic compute needed for high-resolution or multi-dimensional images grows quickly. The authors of this paper propose a physics-driven framework that shifts that burden away from conventional electronics.

Their approach reformulates the imaging inverse problem as an Ising model, a mathematical description of interacting spins whose lowest-energy state encodes the reconstructed image. A non-von Neumann Ising machine then evolves toward that ground state, effectively performing the reconstruction through its own physical dynamics. As a proof of concept, the team used a simulated-annealing-based Ising solver to reconstruct objects for both a single-pixel imaging setup and a lensless diffuser camera, with sampling rates ranging from 1% to 25%.

By adding a foreground-sparsity constraint and a smoothness constraint, the authors report reconstruction quality comparable to or better than mainstream algorithms. They argue that this approach could substantially reduce the electronic computational burden of inverse reconstruction, potentially enabling real-time, low-power computational imaging systems that rely on physical solvers rather than digital processors.