Abstract
Deep learning enables high-quality single-pass reconstruction in computational lensless imaging, but its outputs generally do not satisfy the physical observation model
y
=
A
x
. Existing methods that enforce this constraint, such as iterative optimization and diffusion-based approaches, require repeated computation at inference, limiting their applicability in time-sensitive settings. We propose the output-constraint module (OCM), a lightweight add-on that projects the output of any deep learning reconstruction model onto the measurement-consistent subspace without iterative computation. The projection is derived analytically via Lagrange multipliers, yielding a closed-form expression that requires only matrix-vector products at inference. Applied to lensless image reconstruction, OCM reduces the data-consistency loss (the residual of the observation model) by approximately 3 × over unconstrained deep learning and achieves the best reconstruction quality among measurement-consistent single-pass methods, with inference approximately 9 × faster than iterative optimization (ADMM, 100 iterations). Our analysis indicates that the remaining gap in similarity to the paired lensed-camera images between data-driven and measurement-consistent methods largely reflects the mapping between the two camera domains that data-driven models learn implicitly, rather than a loss of scene information.